grok-1/checkpoint.py

# Copyright 2024 X.AI Corp.
#
# 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.

from __future__ import annotations

import contextlib
import logging
import math
import os
import pickle
import re
import shutil
import sys
import tempfile
from concurrent.futures import ThreadPoolExecutor, wait
from typing import Any, Optional

import jax
import numpy as np
from jax.experimental import multihost_utils

from model import QuantizedWeight8bit

logger = logging.getLogger(__name__)
rank_logger = logging.getLogger("rank")

# Needed for loading the checkpoint with pickle.
sys.modules['__main__'].QuantizedWeight8bit = QuantizedWeight8bit


# Utility functions for file handling and shared memory

@contextlib.contextmanager
def copy_to_shm(file: str):
    """
    Context manager to copy a file to shared memory.

    Args:
        file (str): The path to the file to be copied.

    Yields:
        str: The path to the copied file in shared memory.
    """
    if file.startswith("/dev/shm/"):
        # Nothing to do, the file is already in shared memory.
        yield file
        return

    tmp_dir = "/dev/shm/"
    fd, tmp_path = tempfile.mkstemp(dir=tmp_dir)
    try:
        shutil.copyfile(file, tmp_path)
        yield tmp_path
    finally:
        os.remove(tmp_path)
        os.close(fd)


@contextlib.contextmanager
def copy_from_shm(file: str):
    """
    Context manager to copy a file from shared memory.

    Args:
        file (str): The path to the file to be copied.

    Yields:
        str: The path to the temporary file in shared memory.
    """
    tmp_dir = "/dev/shm/"
    fd, tmp_path = tempfile.mkstemp(dir=tmp_dir)
    try:
        yield tmp_path
        shutil.copyfile(tmp_path, file)
    finally:
        os.remove(tmp_path)
        os.close(fd)


def fast_unpickle(path: str) -> Any:
    """
    Unpickle an object from a file using shared memory for faster loading.

    Args:
        path (str): The path to the file containing the pickled object.

    Returns:
        Any: The unpickled object.
    """
    with copy_to_shm(path) as tmp_path:
        with open(tmp_path, "rb") as f:
            return pickle.load(f)


def fast_pickle(obj: Any, path: str) -> None:
    """
    Pickle an object to a file using shared memory for faster saving.

    Args:
        obj (Any): The object to be pickled.
        path (str): The path to the file where the object will be saved.
    """
    with copy_from_shm(path) as tmp_path:
        with open(tmp_path, "wb") as f:
            pickle.dump(obj, f)


# Tensor loading and path handling

def load_tensors(shaped_arrays, directory, mesh_config, tensor_indices=None):
    """
    Load a set of arrays from files in parallel using a thread pool.

    Args:
        shaped_arrays (list): A list of shaped arrays to be loaded.
        directory (str): The directory containing the tensor files.
        mesh_config (tuple): The mesh configuration.
        tensor_indices (list, optional): The indices of the tensors to load. Defaults to None.

    Returns:
        list: A list of loaded arrays.
    """
    pool = ThreadPoolExecutor(max_workers=32)
    fs = list()
    num_tensors = 0
    num_replicas = 1
    data_model_shards = math.prod(mesh_config)
    if tensor_indices is None:
        iterator = enumerate(shaped_arrays)
    else:
        iterator = zip(tensor_indices, shaped_arrays)
    for i, t in iterator:
        if (i % num_replicas) == ((jax.process_index() // data_model_shards) % num_replicas):
            idx = (
                jax.process_index() // (num_replicas * data_model_shards) * data_model_shards
                + jax.process_index() % data_model_shards
            )
            fs.append(
                pool.submit(fast_unpickle, os.path.join(directory, f"tensor{i:05d}_{idx:03d}"))
            )
            num_tensors += 1
        else:
            fs.append(pool.submit(np.zeros, t.shape, dtype=t.dtype))
    wait(fs)
    return [f.result() for f in fs]


def path_tuple_to_string(path: tuple) -> str:
    """
    Convert a path tuple to a string representation.

    Args:
        path (tuple): The path tuple.

    Returns:
        str: The string representation of the path.
    """
    pieces = []
    for elem in path:
        if isinstance(elem, jax.tree_util.DictKey):
            pieces.append(elem.key)
        elif isinstance(elem, jax.tree_util.GetAttrKey):
            pieces.append(elem.name)
        else:
            assert isinstance(elem, (jax.tree_util.FlattenedIndexKey, jax.tree_util.SequenceKey))
    return "/".join(pieces)


def get_load_path_str(
    init_path_str: str,
    load_rename_rules: Optional[list[tuple[str, str]]] = None,
    load_exclude_rules: Optional[list[str]] = None,
) -> Optional[str]:
    """
    Get the load path string based on the initial path string and renaming/exclusion rules.

    Args:
        init_path_str (str): The initial path string.
        load_rename_rules (list[tuple[str, str]], optional): The renaming rules. Defaults to None.
        load_exclude_rules (list[str], optional): The exclusion rules. Defaults to None.

    Returns:
        Optional[str]: The load path string if not excluded, otherwise None.
    """
    # Exclusion
    if load_exclude_rules is not None:
        for search_pattern in load_exclude_rules:
            if re.search(search_pattern, init_path_str):
                return None

    # Renaming
    load_path_str = init_path_str
    if load_rename_rules is not None:
        for search_pattern, replacement_pattern in load_rename_rules:
            if re.search(search_pattern, load_path_str):
                load_path_str = re.sub(search_pattern, replacement_pattern, load_path_str)
                break

    return load_path_str


def replace_with_load_state(
    init_state: Any,
    load_state: Any,
    load_rename_rules: Optional[list[tuple[str, str]]] = None,
    load_exclude_rules: Optional[list[str]] = None,
    mesh_config: tuple = (1, 1),
) -> Any:
    """
    Replace the initial state with the loaded state based on renaming and exclusion rules.

    Args:
        init_state (Any): The initial state.
        load_state (Any): The loaded state.
        load_rename_rules (list[tuple[str, str]], optional): The renaming rules. Defaults to None.
        load_exclude_rules (list[str], optional): The exclusion rules. Defaults to None.
        mesh_config (tuple, optional): The mesh configuration. Defaults to (1, 1).

    Returns:
        Any: The replaced state.
    """
    flatten_load, _ = jax.tree_util.tree_flatten_with_path(load_state)
    flatten_init, structure_init = jax.tree_util.tree_flatten_with_path(init_state)
    load_map = {path_tuple_to_string(path): tensor for path, tensor in flatten_load}

    replaced = []
    num_replicas = 1
    data_model_shards = math.prod(mesh_config)
    for i, (init_path, tensor) in enumerate(flatten_init):
        init_path_str = path_tuple_to_string(init_path)
        load_path_str = get_load_path_str(init_path_str, load_rename_rules, load_exclude_rules)
        if load_path_str is None:
            rank_logger.info(f"Excluded from restore: {init_path_str}.")
            replaced.append(tensor)
        elif load_path_str in load_map:
            if load_path_str == init_path_str:
                rank_logger.info(f"Restored from ckpt: {init_path_str}.")
            else:
                rank_logger.info(f"Restored from ckpt: {init_path_str} <-- {load_path_str}.")
            replaced.append(load_map[load_path_str])
        else:
            rank_logger.info(f"Not found in ckpt: {init_path_str}.")
            if (i % num_replicas) == ((jax.process_index() // data_model_shards) % num_replicas):
                replaced.append(tensor)
            else:
                replaced.append(np.zeros_like(tensor))

    return jax.tree_util.tree_unflatten(structure_init, replaced)


# Checkpoint restoration

def restore(
    checkpoint_path: str,
    state_shapes: Any,
    mesh,
    between_hosts_config,
    params_only,
    state_sharding,
    init_state: Optional[Any] = None,
) -> Any:
    """
    Restore the state from a checkpoint.

    Args:
        checkpoint_path (str): The path to the checkpoint directory.
        state_shapes (Any): The shapes of the state.
        mesh: The mesh configuration.
        between_hosts_config: The configuration for communication between hosts.
        params_only (bool): Whether to restore only the parameters.
        state_sharding: The sharding specification for the state.
        init_state (Optional[Any], optional): The initial state. Defaults to None.

    Returns:
        Any: The restored state.
    """
    ckpt_path = os.path.join(checkpoint_path, "ckpt-0")

    rank_logger.info("Loading checkpoint at {}".format(ckpt_path))
    ckpt_shapes = state_shapes
    ckpt_shapes_with_path, structure = jax.tree_util.tree_flatten_with_path(ckpt_shapes)

    ckpt_shapes_flat = [elem[1] for elem in ckpt_shapes_with_path]
    loaded_tensors = load_tensors(ckpt_shapes_flat, ckpt_path, between_hosts_config)

    state = jax.tree_util.tree_unflatten(structure, loaded_tensors)

    # Sanity check to give a better error message.
    ckpt_keys = set(state.params.keys())
    code_keys = set(state_sharding.params.keys())

    if ckpt_keys != code_keys and init_state is None:
        missing_in_ckpt = code_keys - ckpt_keys
        missing_locally = ckpt_keys - code_keys
        raise ValueError(
            "Parameters in the code are not matching checkpoint parameters.\n"
            "Params missing in checkpoint: {}\nParams missing in code: {}".format(
                missing_in_ckpt, missing_locally
            )
        )
    state_sharding = jax.tree_util.tree_map(
        lambda x: jax.sharding.PartitionSpec() if x is None else x,
        state_sharding,
        is_leaf=lambda x: x is None,
    )
    state = multihost_utils.host_local_array_to_global_array(state, mesh, state_sharding)
    if params_only:
        state = state.params
    return state