# References:
#   https://github.com/facebookresearch/dino/blob/master/vision_transformer.py
#   https://github.com/rwightman/pytorch-image-models/tree/master/timm/models/vision_transformer.py

import os

from torch import Tensor
from torch import nn

XFORMERS_ENABLED = os.environ.get("XFORMERS_DISABLED") is None
try:
    if XFORMERS_ENABLED:
        from xformers.ops import memory_efficient_attention, unbind

        XFORMERS_AVAILABLE = True
    else:
        raise ImportError
except ImportError:
    XFORMERS_AVAILABLE = False


class Attention(nn.Module):
    def __init__(
        self,
        dim: int,
        num_heads: int = 8,
        qkv_bias: bool = False,
        proj_bias: bool = True,
        attn_drop: float = 0.0,
        proj_drop: float = 0.0,
    ) -> None:
        super().__init__()
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = head_dim**-0.5

        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(dim, dim, bias=proj_bias)
        self.proj_drop = nn.Dropout(proj_drop)

    def forward(self, x: Tensor, return_attn=False) -> Tensor:
        """
        Adapted from https://gitlab.com/ziegleto-machine-learning/dino/-/tree/main/
        """
        B, N, C = x.shape
        qkv = (
            self.qkv(x)
            .reshape(B, N, 3, self.num_heads, C // self.num_heads)
            .permute(2, 0, 3, 1, 4)
        )

        q, k, v = qkv[0] * self.scale, qkv[1], qkv[2]
        attn = q @ k.transpose(-2, -1)

        attn = attn.softmax(dim=-1)
        attn = self.attn_drop(attn)

        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)

        # Adaptation for returing attentions
        if return_attn:
            return attn
        return x


class MemEffAttention(Attention):
    """
    Adapted from https://gitlab.com/ziegleto-machine-learning/dino/-/tree/main/
    """

    def forward(self, x: Tensor, attn_bias=None, return_attn=False) -> Tensor:
        if not XFORMERS_AVAILABLE:
            assert attn_bias is None, "xFormers is required for nested tensors usage"
            # Change this line
            # return super().forward(x)
            # Adaptation for returing attentions
            return super().forward(x, return_attn)

        B, N, C = x.shape
        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads)

        q, k, v = unbind(qkv, 2)

        x = memory_efficient_attention(q, k, v, attn_bias=attn_bias)
        if return_attn:
            # Support for XFORMERS to return attention
            # Adapted from https://github.com/facebookresearch/dinov2/issues/90#issuecomment-1574001076
            attn = x.permute(0, 2, 1, 3) @ v.permute(0, 2, 3, 1)
            return attn
        x = x.reshape([B, N, C])

        x = self.proj(x)
        x = self.proj_drop(x)
        return x


if __name__ == "__main__":
    import torch

    _att = MemEffAttention(dim=32, num_heads=4).to("cuda")
    print(_att(torch.randn(4, 16, 32, device="cuda"), return_attn=True).shape)
    print(_att(torch.randn(4, 16, 32, device="cuda")).shape)