pytorch/torch/distributed/checkpoint/filesystem.py

from abc import ABC, abstractmethod
import queue
import threading
import collections

from dataclasses import dataclass
import os
import dataclasses
import io
import pickle
from typing import Optional, List, Union, Dict, cast

import torch
from torch import Tensor
from torch.futures import Future
from pathlib import Path

from .metadata import (
    Metadata,
    MetadataIndex,
)
from .storage import (
    StorageReader,
    StorageWriter,
    WriteResult,
)

from .planner import (
    LoadItemType,
    LoadPlanner,
    LoadPlan,
    SavePlan,
    SavePlanner,
    ReadItem,
    WriteItem,
    WriteItemType,
)

from .utils import _create_file_view

import torch.distributed as dist
from torch.distributed.checkpoint.checkpointer import Checkpointer
from torch.distributed._shard._utils import narrow_tensor_by_index
from torch._utils import _get_device_module

__all__ = [
    "FileSystemWriter",
    "FileSystemReader",
    "FileSystemCheckpointer"
]


@dataclass
class _StorageInfo:
    """This is the per entry storage info."""

    relative_path: str
    offset: int
    length: int


@dataclass
class _StoragePrefix:
    prefix: str


DEFAULT_SUFFIX = ".distcp"


def _trim(tensor: torch.Tensor) -> torch.Tensor:
    tensor = tensor.detach().cpu()
    if tensor._typed_storage()._size() != tensor.numel():
        tensor = tensor.clone()
    return tensor


def _result_from_write_item(
    item: WriteItem, size_in_bytes, storage_data
) -> WriteResult:
    return WriteResult(
        index=item.index, size_in_bytes=size_in_bytes, storage_data=storage_data
    )


class _TensorLoader(ABC):
    @abstractmethod
    def add(self, size, obj):
        pass

    @abstractmethod
    def start_loading(self):
        pass

    @abstractmethod
    def values(self):
        pass


class _SerialCpuLoader(_TensorLoader):
    def __init__(self, resolve_fun):
        self.resolve_fun = resolve_fun
        self.items = []

    def add(self, size, obj):
        self.items.append((size, obj))

    def start_loading(self):
        pass

    def values(self):
        for _, obj in self.items:
            tensor = self.resolve_fun(obj).detach()
            tensor = tensor.cpu()
            if tensor.storage().size() != tensor.numel():
                tensor = tensor.clone()
            yield (
                tensor,
                obj,
            )


class _OverlappingCpuLoader(_TensorLoader):
    def __init__(self, resolve_fun, stream=None, inflight_threshhold=1_000_000):
        self.resolve_fun = resolve_fun
        self.items = []
        self.inflight_threshhold = inflight_threshhold
        self.in_flight_data = 0
        self.current_items: collections.deque = collections.deque()
        self.idx = 0
        self.started = False
        self.device_type = stream.device_type if stream else torch.device("cuda").type
        self.device_module = _get_device_module(self.device_type)
        self.stream = stream or self.device_module.current_stream()
        if self.stream != self.device_module.current_stream():
            self.stream.wait_stream(self.device_module.current_stream())

    @property
    def _done(self):
        return self.idx >= len(self.items)

    def _drain(self):
        drained = []
        if self.in_flight_data >= self.inflight_threshhold:
            self.stream.synchronize()
        while self.in_flight_data >= self.inflight_threshhold:
            val = self.current_items.popleft()
            self.in_flight_data -= val[0].numel() * val[0].element_size()
            drained.append(val)
        return drained

    def _refill(self):
        with self.device_module.stream(self.stream):
            while (
                not self._done
                and self.in_flight_data < self.inflight_threshhold
            ):
                _, obj = self.items[self.idx]
                self.idx += 1
                tensor = self.resolve_fun(obj).detach()
                if tensor.device.type == self.device_type:
                    tensor = tensor.to(device="cpu", non_blocking=True)
                elif tensor.device == torch.device("cpu"):
                    if tensor.storage().size() != tensor.numel():
                        # this forces the tensor to be both contiguous and with minimal storage
                        tensor = tensor.clone()

                self.current_items.append(
                    (
                        tensor,
                        obj,
                    )
                )
                self.in_flight_data += tensor.numel() * tensor.element_size()

    def _finish(self):
        assert self._done
        if len(self.current_items) > 0:
            self.stream.synchronize()
        return self.current_items

    def add(self, size, obj):
        if self.started:
            raise RuntimeError("cannot add items after loading started")
        self.items.append((size, obj))

    def start_loading(self):
        if self.started:
            return
        self.started = True
        self.items.sort(key=lambda x: x[0])
        self._refill()

    def values(self):
        self.start_loading()
        while not self._done:
            drained = self._drain()
            self._refill()
            yield from drained

        yield from self._finish()


def _item_size(item: WriteItem) -> int:
    size = 1
    assert item.tensor_data is not None
    # can't use math.prod as PT needs to support older python
    for s in item.tensor_data.size:
        size *= s

    dtype = item.tensor_data.properties.dtype
    return size * torch._utils._element_size(dtype)


def _split_by_size_and_type(
    bins, items: List[WriteItem]
) -> List[List[WriteItem]]:
    if bins == 1:
        return [items]

    bytes_w = [wi for wi in items if wi.type == WriteItemType.BYTE_IO]
    tensor_w = [wi for wi in items if wi.type != WriteItemType.BYTE_IO]

    buckets: List[List[WriteItem]] = [[] for _ in range(bins)]
    bucket_sizes = [0 for _ in range(bins)]

    tensor_w.sort(key=_item_size, reverse=True)

    for i, wi in enumerate(bytes_w):
        buckets[i % bins].append(wi)

    for wi in tensor_w:
        # TODO replace with headq
        idx = min(enumerate(bucket_sizes), key=lambda x: x[1])[0]
        buckets[idx].append(wi)
        bucket_sizes[idx] += _item_size(wi)

    return buckets


def _write_item(stream, data, write_item, storage_key):
    offset = stream.tell()

    if write_item.type == WriteItemType.BYTE_IO:
        assert isinstance(data, io.BytesIO)
        stream.write(data.getbuffer())
    else:
        assert isinstance(data, torch.Tensor)
        assert data.device == torch.device("cpu")
        torch.save(data, stream)
    length = stream.tell() - offset

    return _result_from_write_item(
        write_item, length, _StorageInfo(storage_key, offset, length)
    )


def _write_files_from_queue(
    file_queue: queue.Queue,
    result_queue: queue.Queue,
    planner: SavePlanner,
    inflight_threshhold: int,
    use_fsync: bool,
):
    try:
        while True:
            file_name, storage_key, write_items = file_queue.get_nowait()
            loader: _TensorLoader

            if torch.cuda.is_available() and inflight_threshhold > 0:
                loader = _OverlappingCpuLoader(
                    planner.resolve_data,
                    inflight_threshhold=inflight_threshhold,
                )
            else:
                loader = _SerialCpuLoader(
                    planner.resolve_data,
                )

            tensor_w = [
                wi for wi in write_items if wi.type != WriteItemType.BYTE_IO
            ]
            for write_item in tensor_w:
                loader.add(_item_size(write_item), write_item)
            loader.start_loading()

            bytes_w = [
                wi for wi in write_items if wi.type == WriteItemType.BYTE_IO
            ]
            write_results = []

            with file_name.open("wb") as stream:
                for write_item in bytes_w:
                    data = planner.resolve_data(write_item)
                    write_results.append(
                        _write_item(stream, data, write_item, storage_key)
                    )

                for tensor, write_item in loader.values():
                    assert tensor.is_cpu
                    write_results.append(
                        _write_item(stream, tensor, write_item, storage_key)
                    )

                if use_fsync:
                    os.fsync(stream.fileno())
            result_queue.put(write_results)
    except queue.Empty:
        pass


class FileSystemWriter(StorageWriter):
    """
    Basic implementation of StorageWriter using file IO.

    This implementation makes the following assumptions and simplifications:

    * The checkpoint path is an empty or non-existing directory.
    * File creation is atomic

    The checkpoint consist of one file per write request plus
    a `.metadata` file with the serialized metadata.

    """

    def __init__(
        self,
        path: Union[str, os.PathLike],
        single_file_per_rank: bool = True,
        sync_files: bool = True,
        thread_count: int = 1,
        per_thread_copy_ahead: int = 10_000_000,
    ) -> None:
        """
        Initialize the writer pointing to `path`.

        Args:
            path: directory where the checkpoint will be written to.
            single_file_per_rank: Produce one file per rank instead of one file per tensor/blob. Default to True.
            sync_files : force files to be synced to permanent storage. Default to True.
            thread_count: Number of IO threads to use to write. Default to 1.
            per_thread_copy_ahead: How many bytes to copy from the GPU ahead of saving then. Default 10Mb.

        N. B. If sync_files is disabled, there's no guarantee that the checkpoint will be consistent in the case of a failure.
        """
        super().__init__()
        if not isinstance(path, Path):
            path = Path(path)
        self.path = path
        self.single_file_per_rank = single_file_per_rank
        self.sync_files = sync_files
        self.thread_count = thread_count
        self.per_thread_copy_ahead = per_thread_copy_ahead

    def set_up_storage_writer(self, is_coordinator: bool) -> None:
        pass

    def prepare_local_plan(self, plan: SavePlan) -> SavePlan:
        self.path.mkdir(parents=True, exist_ok=True)
        return plan

    def prepare_global_plan(
        self, global_plan: List[SavePlan]
    ) -> List[SavePlan]:
        new_plans = [
            dataclasses.replace(plan, storage_data=_StoragePrefix(f"__{i}_"))
            for i, plan in enumerate(global_plan)
        ]
        return new_plans

    def write_data(
        self,
        plan: SavePlan,
        planner: SavePlanner,
    ) -> Future[List[WriteResult]]:
        storage_plan: _StoragePrefix = plan.storage_data
        file_count = 0

        def gen_file():
            nonlocal file_count
            file_name = f"{storage_plan.prefix}{file_count}{DEFAULT_SUFFIX}"
            file_count += 1
            return file_name

        file_queue: queue.Queue = queue.Queue()
        if self.single_file_per_rank:
            for bucket in _split_by_size_and_type(
                self.thread_count, plan.items
            ):
                file_name = gen_file()
                file_queue.put((self.path / file_name, file_name, bucket))
        else:
            for item in plan.items:
                file_name = gen_file()
                file_queue.put((self.path / file_name, file_name, [item]))

        result_queue: queue.Queue = queue.Queue()

        threads = []
        for _ in range(1, self.thread_count):
            t = threading.Thread(
                target=_write_files_from_queue,
                args=(
                    file_queue,
                    result_queue,
                    planner,
                    self.per_thread_copy_ahead,
                    self.sync_files,
                ),
            )
            t.start()
            threads.append(t)

        _write_files_from_queue(
            file_queue=file_queue,
            result_queue=result_queue,
            planner=planner,
            inflight_threshhold=self.per_thread_copy_ahead,
            use_fsync=self.sync_files,
        )

        for t in threads:
            t.join()

        res = []
        try:
            while True:
                res += result_queue.get_nowait()
        except queue.Empty:
            pass

            fut: Future[List[WriteResult]] = Future()
            fut.set_result(res)
            return fut

    def finish(
        self, metadata: Metadata, results: List[List[WriteResult]]
    ) -> None:
        storage_md = dict()
        for wr_list in results:
            storage_md.update({wr.index: wr.storage_data for wr in wr_list})
        metadata.storage_data = storage_md
        with (self.path / ".metadata.tmp").open("wb") as metadata_file:
            pickle.dump(metadata, metadata_file)
            if self.sync_files:
                os.fsync(metadata_file.fileno())

        (self.path / ".metadata.tmp").rename(self.path / ".metadata")


class FileSystemReader(StorageReader):
    def __init__(self, path: Union[str, os.PathLike]) -> None:
        super().__init__()
        if not isinstance(path, Path):
            path = Path(path)
        self.path = path
        self.storage_data: Dict[MetadataIndex, _StorageInfo] = dict()

    def _slice_file(self, file, sinfo: _StorageInfo):
        return _create_file_view(file, sinfo.offset, sinfo.length)

    def read_data(self, plan: LoadPlan, planner: LoadPlanner) -> Future[None]:
        # group requests by file
        per_file: Dict[str, List[ReadItem]] = dict()
        for read_item in plan.items:
            item_md = self.storage_data[read_item.storage_index]
            path = item_md.relative_path
            per_file.setdefault(path, []).append(read_item)

        for relative_path, reqs in per_file.items():
            with (self.path / relative_path).open("rb") as file:
                # TODO sort by offset and cache the reading
                for req in reqs:
                    item_md = self.storage_data[req.storage_index]
                    file_slice = self._slice_file(file, item_md)
                    if req.type == LoadItemType.BYTE_IO:
                        bytes = io.BytesIO(file_slice.read(item_md.length))
                        bytes.seek(0)
                        planner.load_bytes(req, bytes)
                    else:
                        tensor = cast(
                            Tensor, torch.load(file_slice, map_location="cpu")
                        )
                        tensor = narrow_tensor_by_index(
                            tensor, req.storage_offsets, req.lengths
                        )
                        target_tensor = planner.resolve_tensor(req).detach()

                        assert (
                            target_tensor.size() == tensor.size()
                        ), f"req {req.storage_index} mismatch sizes {target_tensor.size()} vs {tensor.size()}"
                        target_tensor.copy_(tensor)
                        planner.commit_tensor(req, target_tensor)

        fut: Future = Future()
        fut.set_result(None)
        return fut

    # Implementing the abstract function in StorageReader
    def read_metadata(self) -> Metadata:
        with (self.path / ".metadata").open("rb") as metadata_file:
            return pickle.load(metadata_file)

    def set_up_storage_reader(self, metadata: Metadata, is_coordinator: bool) -> None:
        self.storage_data = metadata.storage_data
        assert self.storage_data is not None

    def prepare_local_plan(self, plan: LoadPlan) -> LoadPlan:
        return plan

    def prepare_global_plan(
        self, global_plan: List[LoadPlan]
    ) -> List[LoadPlan]:
        return global_plan

class FileSystemCheckpointer(Checkpointer):
    """An implementation of :py:class:`torch.distributed.checkpoint.checkpointer.Checkpointer`
    for the file system. Wraps the creation and usage of ``FileSystemWriter`` and ``FileSystemReader``.
    """

    def __init__(
        self,
        path: Union[str, os.PathLike],
        *,
        single_file_per_rank: bool = True,
        sync_files: bool = True,
        thread_count: int = 1,
        per_thread_copy_ahead: int = 10_000_000,
        process_group: Optional[dist.ProcessGroup] = None,
        coordinator_rank: int = 0,
        no_dist: bool = False,
        load_planner: Optional[LoadPlanner] = None,
        save_planner: Optional[SavePlanner] = None,
    ):
        """Initializes Checkpointing defualts, including ``FileSystemWriter`` and ``FileSystemReader``

        Args:
            path: The directory to store/load checkpoints.
            single_file_per_rank: Produce one file per rank instead of one file per tensor/blob. Default to True.
            sync_files: force files to be synced to permanent storage. Default to True.
            thread_count: Number of IO threads to use to write. Default to 1.
            per_thread_copy_ahead: How many bytes to copy from the GPU ahead of saving then. Default 10Mb.
            process_group: ProcessGroup to be used for cross-rank synchronization.
            coordinator_rank: Rank to use to coordinate the checkpoint. rank0 is used by default.
            no_dist: If ``True``, distributed checkpoint will not load in SPMD style. (Default: ``False``)
            loader_planner: Instance of LoadPlanner to use when loading.
            save_planner: Instance of SavePlanner to use when saving.
        """

        storage_writer = FileSystemWriter(
            path,
            single_file_per_rank,
            sync_files,
            thread_count,
            per_thread_copy_ahead
        )
        storage_reader = FileSystemReader(path)

        super().__init__(
            storage_writer,
            storage_reader,
            process_group=process_group,
            coordinator_rank=coordinator_rank,
            no_dist=no_dist,
            load_planner=load_planner,
            save_planner=save_planner
        )