reproducibility.py

# Copyright 2022 MosaicML Composer authors
# SPDX-License-Identifier: Apache-2.0

"""Helper utilities for configuring deterministic training to ensure reproducibility.

.. note::

    For deterministic model initialization, :func:`~.seed_all` and/or
    :func:`~.configure_deterministic_mode` should be
    invoked before creating and initializing a model, before creating the :class:`~.Trainer`.
    For example:

    .. testsetup::

        import functools
        import torch.nn
        import warnings

        warnings.filterwarnings(action="ignore", message="Deterministic mode is activated.")

        MyModel = Model

    .. doctest::

        >>> import torch.nn
        >>> from composer.utils import reproducibility
        >>> reproducibility.configure_deterministic_mode()
        >>> reproducibility.seed_all(42)
        >>> model = MyModel()
        >>> def init_weights(m):
        ...     if isinstance(m, torch.nn.Linear):
        ...         torch.nn.init.xavier_uniform(m.weight)
        >>> # model will now be deterministically initialized, since the seed is set.
        >>> init_weights(model)
        >>> trainer = Trainer(model=model, seed=42)

    Note that the seed must also be passed to the Trainer, otherwise the Trainer
    would generate a random seed based on the timestamp (see :func:`~.get_random_seed`).

    .. testcleanup::

        warnings.resetwarnings()

Attributes:
    MAX_SEED (int): The maximum allowed seed, which is :math:`2^{32} - 1`.
"""

from __future__ import annotations

import logging
import os
import random
import textwrap
import time
import warnings
from typing import Any, Dict, List

import numpy as np
import torch
import torch.backends.cudnn

from composer.utils import dist

__all__ = [
    'configure_deterministic_mode',
    'get_random_seed',
    'seed_all',
    'get_rng_state',
    'load_rng_state',
    'MAX_SEED',
]

log = logging.getLogger(__name__)

# seeds must be 32-bit unsigned integers
MAX_SEED = 2**32 - 1


def configure_deterministic_mode():
    """Configure PyTorch deterministic mode.

    .. note::

        When using the :class:`.Trainer`, you can use the ``deterministic_mode`` flag
        instead of invoking this function directly.
        For example:

        .. testsetup::

            import warnings

            warnings.filterwarnings(action="ignore", message="Deterministic mode is activated.")

        .. doctest::

            >>> trainer = Trainer(deterministic_mode=True)

        .. testcleanup::

            warnings.resetwarnings()

        However, to configure deterministic mode for operations before the trainer is initialized, manually invoke this
        function at the beginning of your training script.

    .. note::

        When training on a GPU, this function must be invoked before any CUDA operations.

    .. note::

        Deterministic mode degrades performance. Do not use outside of testing and debugging.
    """
    torch.use_deterministic_algorithms(True)
    torch.backends.cudnn.benchmark = False
    torch.backends.cudnn.deterministic = True
    # See https://pytorch.org/docs/stable/generated/torch.use_deterministic_algorithms.html
    # and https://docs.nvidia.com/cuda/cublas/index.html#cublasApi_reproducibility
    os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8'
    warnings.warn('Deterministic mode is activated. This will negatively impact performance.', category=UserWarning)


def get_random_seed() -> int:
    """Get a randomly created seed to use for seeding rng objects.

    .. warning::

        This random seed is NOT cryptographically secure.

    Returns:
        int: A random seed.
    """
    rng = random.Random(int(time.time_ns()))  # get a new RNG does not respect the current seed
    seed = rng.randint(0, MAX_SEED)
    assert seed >= 0 and seed <= MAX_SEED, 'seed should be on this range'
    return seed


def seed_all(seed: int):
    """Seed all rng objects.

    .. note::

        When using the :class:`.Trainer`, you can use the ``seed`` parameter
        instead of invoking this function directly.
        For example:

        .. doctest::

            >>> trainer = Trainer(seed=42)

        However, to configure the random seed for operations before the trainer is initialized, manually invoke this
        function at the beginning of your training script.

    Args:
        seed (int): The random seed
    """
    if seed < 0 or seed > MAX_SEED:
        raise ValueError(f'Seed {seed} is invalid. It must be on [0; 2^32 - 1]')
    log.info('Setting seed to %d', seed)
    random.seed(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)
    # torch.manual_seed may call manual_seed_all but calling it again here
    # to make sure it gets called at least once
    torch.cuda.manual_seed_all(seed)


def get_rng_state() -> List[Dict[str, Any]]:
    """The state of the RNG objects.

    Returns:
        List[Dict[str, Any]]: A list of RNG State Dicts, indexed by global rank.
    """
    rng_state = {
        'python': random.getstate(),
        'numpy': np.random.get_state(),
        'torch': torch.random.get_rng_state(),
    }
    if torch.cuda.is_available() and torch.cuda.is_initialized():
        # This will not be compatible with model parallelism
        rng_state['cuda'] = torch.cuda.get_rng_state()

    return dist.all_gather_object(rng_state)


def load_rng_state(rng_state_dicts: List[Dict[str, Any]]):
    """Restore the RNG state.

    Args:
        rng_state_dicts (List[Dict[str, Any]]): The list of RNG state dicts to restore,
            as returned by :func:`get_rng_state`.
    """
    if dist.get_world_size() > len(rng_state_dicts):
        warnings.warn(
            textwrap.dedent(f"""\
                The current world size ({dist.get_world_size()} is greater than the number of RNG state(s) serialized
                ({len(rng_state_dicts)}). Only the first {len(rng_state_dicts)} rank(s) will have their RNG restored.
                """))
    if dist.get_world_size() < len(rng_state_dicts):
        warnings.warn(
            textwrap.dedent(f"""\
            The current world size ({dist.get_world_size()} is less than the number of RNG state(s) serialized
            ({len(rng_state_dicts)}). Only the first {dist.get_world_size()} RNG state(s) will be consumed;
            the remaining will be ignored."""))

    if dist.get_global_rank() < len(rng_state_dicts):
        rng_state_dict = rng_state_dicts[dist.get_global_rank()]
        torch.set_rng_state(rng_state_dict['torch'])
        random.setstate(rng_state_dict['python'])
        np.random.set_state(rng_state_dict['numpy'])

        is_cuda_available = torch.cuda.is_available() and torch.cuda.is_initialized()
        has_cuda_rng_state = 'cuda' in rng_state_dict

        log.debug('Restoring the RNG state')

        if is_cuda_available and has_cuda_rng_state:
            torch.cuda.set_rng_state(rng_state_dict['cuda'])

        if is_cuda_available and not has_cuda_rng_state:
            warnings.warn(
                textwrap.dedent(f"""\
                The checkpoint did not include the CUDA RNG state. The CUDA RNG will have a
                non-deterministic state."""))
        if not is_cuda_available and has_cuda_rng_state:
            warnings.warn(
                textwrap.dedent(f"""\
                The checkpoint included CUDA RNG state, but CUDA is not being used.
                As such, the CUDA RNG state will be ignored."""))