ColossalAI/colossalai/tensor/colo_tensor.py

from .op_wrapper import _COLOSSAL_OPS

import torch
from typing import Tuple, Optional
from numpy import product
from colossalai.core import global_context as gpc
from colossalai.context import ParallelMode
from colossalai.nn.layer.utils import divide
from colossalai.utils.cuda import get_current_device


class ColoTensor(object):
    """ Data Structure for Tensor in Colossal-AI
    1. It contains a torch.Tensor as an attribute.
    2. It supports lazy init the tensor's payload.
    3. It can hijack the torch functions which using ColoTensors as args to our customized functions.
    4. It supports distributing the tensor's payload to the shards among processes. (TODO)
    """

    def __new__(cls, *args, **kwargs):
        return super(ColoTensor, cls).__new__(cls)

    def __init__(
            self,
            *size: Tuple[int],
            dtype=None,
            requires_grad=False,
            pin_memory=False,
            device=None,
            torch_tensor=torch.empty(0),
            shard_spec: str = None,
    ):
        self._size = size
        self._dtype = dtype
        self._requires_grad = requires_grad
        self._pin_memory = pin_memory
        self._device = device
        self._torch_tensor = torch_tensor
        self._shard_spec = shard_spec

    @property
    def shard_spec(self) -> Optional[str]:
        return self._shard_spec

    @property
    def data(self):
        return self._torch_tensor.data

    @property
    def grad(self):
        return self._torch_tensor.grad

    @property
    def size(self):
        return self._size

    @property
    def shape(self):
        return torch.Size(self._size)

    @property
    def device(self):
        return self._torch_tensor.device

    def size(self, dim=None):
        if dim is None:
            return self.shape
        return self._size[dim]

    def dim(self):
        return len(self._size)

    def normal_(self, mean=0., std=1.):
        torch_tensor = self.torch_tensor()
        return torch_tensor.normal_(mean=mean, std=std)

    def numel(self):
        return product(self._size)

    @staticmethod
    def init_from_torch_tensor(tensor: torch.Tensor, save_payload=True) -> 'ColoTensor':
        colo_t = ColoTensor(*tensor.size(),
                            dtype=tensor.dtype,
                            requires_grad=tensor.requires_grad,
                            pin_memory=tensor.is_pinned(),
                            device=tensor.device,
                            torch_tensor=tensor if save_payload else torch.empty(0))
        return colo_t

    def del_torch_tensor(self, save_shape=False) -> None:
        """
        delete the payload of the torch tensor.

        Args:
            save_shape (bool, optional): if saving the shape of the torch_tensor. 
            If saving the shape, the size of self._torch_tensor is inconsist with the self._size.
            Defaults to False.
        """
        if not save_shape:
            self._size = (0,)
        self._torch_tensor = torch.empty((0,), device=self._device, dtype=self._dtype)

    def torch_tensor(self) -> torch.Tensor:
        if self._torch_tensor.numel() == 0:
            self._torch_tensor = torch.empty(*self._size,
                                             dtype=self._dtype,
                                             pin_memory=self._pin_memory,
                                             requires_grad=self._requires_grad,
                                             device=self._device)
        return self._torch_tensor

    def set_spec(self, spec: str, lazy_shard: bool = False) -> None:
        self._shard_spec = spec
        if lazy_shard == False:
            self._shard()

    def _shard(self):
        assert self._shard_spec is not None, 'You should call set_spec() before _shard() ColoTensor.'
        if self._shard_spec == "1Drow":    # TODO It actually represents the sharding layout for Linear-1Drow-weight, but we make it simpler now.
            num_partition = gpc.get_world_size(ParallelMode.TENSOR)
            local_rank = gpc.get_local_rank(ParallelMode.TENSOR)
            dim = -1
            chunk_size = divide(self._size[dim], num_partition)
            device = get_current_device()
            # Reshape to get shard for this rank and we don't want autograd
            # recording here for the narrow op and 'local_shard' should be a
            # leaf variable in the autograd graph.
            self._torch_tensor = self._torch_tensor.narrow(dim, local_rank * chunk_size, chunk_size).detach(
            ).contiguous()    # TODO Shall we clone() here since detach() will point to the old tensor?
            self._torch_tensor.requires_grad = self._requires_grad
            self._size = self._torch_tensor.size()
            self._device = device    # TODO A `fake` device now because torch_tensor.device always = cpu

    @classmethod
    def __torch_function__(cls, func, types, args=(), kwargs=None):
        global _COLOSSAL_OPS
        if func in _COLOSSAL_OPS:
            for arg in args:
                if isinstance(arg, ColoTensor):
                    return _COLOSSAL_OPS[func](types, args, kwargs, None)

            for kwarg in kwargs.values():
                if isinstance(kwarg, ColoTensor):
                    return _COLOSSAL_OPS[func](types, args, kwargs, None)
        else:
            # If we have not hijact the function, convert the ColoTensors in args and kwargs to torch tensors.
            args = [arg.torch_tensor() if isinstance(arg, ColoTensor) else arg for arg in args]
            if kwargs is None:
                kwargs = {}

            kwargs = {k: v.torch_tensor() if isinstance(v, ColoTensor) else v for k, v in kwargs.items()}
            return func(*args, **kwargs)

    def backward(self, retain_graph: bool = False):
        self._torch_tensor.backward(retain_graph=retain_graph)