xtuner/parallel/sequence/comm.py

# Copyright (c) OpenMMLab. All rights reserved.
from typing import Any, Tuple

import torch
import torch.distributed as dist
from torch import Tensor


def _all_to_all(
    input: Tensor,
    world_size: int,
    group: dist.ProcessGroup,
    scatter_dim: int,
    gather_dim: int,
):
    input_list = [
        t.contiguous()
        for t in torch.tensor_split(input, world_size, scatter_dim)
    ]
    output_list = [torch.empty_like(input_list[0]) for _ in range(world_size)]
    dist.all_to_all(output_list, input_list, group=group)
    return torch.cat(output_list, dim=gather_dim).contiguous()


class _AllToAll(torch.autograd.Function):
    """All-to-all communication.

    Args:
        input: Input tensor
        sp_group: Sequence parallel process group
        scatter_dim: Scatter dimension
        gather_dim: Gather dimension
    """

    @staticmethod
    def forward(ctx: Any, input: Tensor, sp_group: dist.ProcessGroup,
                scatter_dim: int, gather_dim: int):
        ctx.sp_group = sp_group
        ctx.scatter_dim = scatter_dim
        ctx.gather_dim = gather_dim
        ctx.world_size = dist.get_world_size(sp_group)
        output = _all_to_all(input, ctx.world_size, sp_group, scatter_dim,
                             gather_dim)
        return output

    @staticmethod
    def backward(ctx: Any, grad_output: Tensor) -> Tuple:
        grad_output = _all_to_all(
            grad_output,
            ctx.world_size,
            ctx.sp_group,
            ctx.gather_dim,
            ctx.scatter_dim,
        )
        return (
            grad_output,
            None,
            None,
            None,
        )


def all_to_all(
    input: Tensor,
    sp_group: dist.ProcessGroup,
    scatter_dim: int = 2,
    gather_dim: int = 1,
):
    """Convenience function to apply the all-to-all operation with scatter and
    gather dimensions.

    Notes:
        We have wrapped the `torch.distributed.all_to_all` function to
        enable automatic differentiation of the all-to-all operation.

    Args:
        input: The input tensor for which all-to-all communication is performed
        sp_group: The sequence parallel process group.
        scatter_dim: The dimension along which the input tensor is scattered
            (default: 2).
        gather_dim: The dimension along which the output tensor is gathered
            (default: 1).

    Returns:
        The output tensor after the all-to-all communication.
    """
    return _AllToAll.apply(input, sp_group, scatter_dim, gather_dim)


def split_for_sequence_parallel(input, dim: int, sp_group: dist.ProcessGroup):
    """Splits the input tensor along a given dimension for sequence parallel.

    Args:
        input: The input tensor to be split.
        dim: The dimension along which the tensor should be split.
        sp_group: The sequence parallel process group.

    Returns:
        The split tensor corresponding to the current rank's chunk.
    """
    world_size = dist.get_world_size(sp_group)
    if world_size == 1:
        return input

    rank = dist.get_rank(sp_group)
    dim_size = input.size(dim)
    assert dim_size % world_size == 0, (
        f'The dimension to split ({dim_size}) is not a multiple of '
        f'world size ({world_size}), cannot split tensor evenly')

    tensor_list = torch.split(input, dim_size // world_size, dim=dim)
    output = tensor_list[rank].contiguous()

    return output


def gather_for_sequence_parallel(input, dim: int, sp_group: dist.ProcessGroup):
    """Gathers the input tensor along a given dimension for sequence parallel.

    Args:
        input: The input tensor to be gathered.
        dim: The dimension along which the tensor should be gathered.
        sp_group: The sequence parallel process group.

    Returns:
        The gathered tensor concatenated along the specified dimension.
    """
    input = input.contiguous()
    world_size = dist.get_world_size(sp_group)
    dist.get_rank(sp_group)

    if world_size == 1:
        return input

    tensor_list = [torch.empty_like(input) for _ in range(world_size)]
    assert input.device.type == 'cuda'
    dist.all_gather(tensor_list, input, group=sp_group)

    output = torch.cat(tensor_list, dim=dim).contiguous()

    return output


class _GatherForwardSplitBackward(torch.autograd.Function):
    """Gather the input during forward.

    Scale and split the grad and keep only the corresponding chuck to the rank
    during backward.
    """

    @staticmethod
    def forward(ctx, input, dim, sp_group, grad_scale):
        ctx.dim = dim
        ctx.sp_group = sp_group
        ctx.grad_scale = grad_scale
        return gather_for_sequence_parallel(input, dim, sp_group)

    @staticmethod
    def backward(ctx, grad_output):
        if ctx.grad_scale == 'up':
            grad_output = grad_output * dist.get_world_size(ctx.sp_group)
        elif ctx.grad_scale == 'down':
            grad_output = grad_output / dist.get_world_size(ctx.sp_group)

        return (split_for_sequence_parallel(grad_output, ctx.dim,
                                            ctx.sp_group), None, None, None)


class _SplitForwardGatherBackward(torch.autograd.Function):
    """Split the input and keep only the corresponding chuck to the rank during
    forward.

    Scale and gather the grad during backward.
    """

    @staticmethod
    def forward(ctx, input, dim, sp_group, grad_scale):
        ctx.dim = dim
        ctx.sp_group = sp_group
        ctx.grad_scale = grad_scale
        return split_for_sequence_parallel(input, dim, sp_group)

    @staticmethod
    def backward(ctx, grad_output):
        if ctx.grad_scale == 'up':
            grad_output = grad_output * dist.get_world_size(ctx.sp_group)
        elif ctx.grad_scale == 'down':
            grad_output = grad_output / dist.get_world_size(ctx.sp_group)
        return (gather_for_sequence_parallel(grad_output, ctx.dim,
                                             ctx.sp_group), None, None, None)


def split_forward_gather_backward(input, dim, sp_group, grad_scale=None):
    """Split tensors according to the sp rank during forward propagation and
    gather the grad from the whole sp group during backward propagation.

    1. When do we need this? input.requires_grad = True

    2. Why we need grad scale?

    We have to scale down the grads as `gather_forward_split_backward` scales
    up the grads.
    """
    return _SplitForwardGatherBackward.apply(input, dim, sp_group, grad_scale)


def gather_forward_split_backward(input, dim, sp_group, grad_scale=None):
    """Gather tensors from the whole sp group during forward propagation and
    split the grad according to the sp rank during backward propagation.

    1. When do we need this?

    When sp is greater than 1, we need to slice the input `x` along
    sequence length dimension before it is passed into the model and get
    `sub_seq_x`. We then pass `sub_seq_x` into model and get output
    `sub_seq_out`. If the loss calculation process needs to use the complete
    output, we have to gather the `sub_seq_out` in all sp ranks during forward
    propagation and split the grad during backward propagation.

    2. Why we need grad scale?
    Here is a simple case.

    -------- SP 1 -----------
    Suppose here is a toy model with only one linear module
    (in_features = 2, out_features = 1) and the input x has shape(2, 2).
    Y = [[y1], = [[w11x11 + w21x12], = [[x11, x12], dot [[w11],
         [y2]]    [w11x21 + w21x22]]    [x21, x22]]      [w21]]
    z = mean(Y) = (y1 + y2) / 2
    Here is the partial derivative of z with respect to w11:
    ∂z / ∂w11 = ∂z / ∂y1 * ∂y1 / ∂w11 + ∂z / ∂y2 * ∂y2 / ∂w11
              = 1/2 * x11 + 1/2 * x21 = (x11 + x21) / 2

    -------- SP 2 -----------
    When sequence parallel world size is set to 2, we will split the input x
    and scatter them to the two rank in the same sequence parallel group.
    ```Step 1
    Y_rank0 = [[y1]] = [[w11x11 + w21x12]] = [[x11, x12]] dot [[w11, w21]]^T
    Y_rank1 = [[y2]] = [[w11x21 + w21x22]] = [[x21, x22]] dot [[w11, w21]]^T
    ```

    Then, we have to gather them:
    ```Step 2
    Y_rank0 = [[y1],
               detach([y2])]
    Y_rank1 = [detach([y1]),
               [y2]]
    ```
    Note that y2 in Y_rank0 does not have grad, neither does y1 in Y_rank1.

    Similarly, we calculate the loss in each rank:
    ```Step 3
    z_rank0 = mean(Y_rank0) = (y1 + detach(y2)) / 2
    z_rank1 = mean(Y_rank1) = (detach(y1) + y2) / 2
    ```
    So the partial derivative of loss_rank0 with respect to w11:
    ```∂z / ∂w11 = ∂z / ∂y1 * ∂y1 / ∂w11 = x11 / 2```
    The same for rank1:
    ```∂z / ∂w11 = ∂z / ∂y2 * ∂y2 / ∂w11 = x21 / 2```

    Finally, we need to all_reduce them:
    ```Step 4
    In both rank:
    ∂z / ∂w11 = (x11 / 2 + x21 / 2) / 2 = (x11 + x21) / 4
    ```

    In SP2, the gradient of each param is only half of that in SP1.
    So we should scale up the grad during the backward process in Step 2.
    """  # noqa: E501
    return _GatherForwardSplitBackward.apply(input, dim, sp_group, grad_scale)