import torch
import torch.nn as nn
import torch.utils.checkpoint

from videosys.models.modules.normalization import LlamaRMSNorm


class Attention(nn.Module):
    def __init__(
        self,
        dim: int,
        num_heads: int = 8,
        qkv_bias: bool = False,
        qk_norm: bool = False,
        attn_drop: float = 0.0,
        proj_drop: float = 0.0,
        norm_layer: nn.Module = LlamaRMSNorm,
        enable_flash_attn: bool = False,
        rope=None,
        qk_norm_legacy: bool = False,
    ) -> None:
        super().__init__()
        assert dim % num_heads == 0, "dim should be divisible by num_heads"
        self.dim = dim
        self.num_heads = num_heads
        self.head_dim = dim // num_heads
        self.scale = self.head_dim**-0.5
        self.enable_flash_attn = enable_flash_attn

        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
        self.q_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
        self.k_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
        self.qk_norm_legacy = qk_norm_legacy
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(dim, dim)
        self.proj_drop = nn.Dropout(proj_drop)

        self.rope = False
        if rope is not None:
            self.rope = True
            self.rotary_emb = rope

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        B, N, C = x.shape
        # flash attn is not memory efficient for small sequences, this is empirical
        enable_flash_attn = self.enable_flash_attn and (N > B)
        qkv = self.qkv(x)
        qkv_shape = (B, N, 3, self.num_heads, self.head_dim)

        qkv = qkv.view(qkv_shape).permute(2, 0, 3, 1, 4)
        q, k, v = qkv.unbind(0)
        if self.qk_norm_legacy:
            # WARNING: this may be a bug
            if self.rope:
                q = self.rotary_emb(q)
                k = self.rotary_emb(k)
            q, k = self.q_norm(q), self.k_norm(k)
        else:
            q, k = self.q_norm(q), self.k_norm(k)
            if self.rope:
                q = self.rotary_emb(q)
                k = self.rotary_emb(k)

        if enable_flash_attn:
            from flash_attn import flash_attn_func

            # (B, #heads, N, #dim) -> (B, N, #heads, #dim)
            q = q.permute(0, 2, 1, 3)
            k = k.permute(0, 2, 1, 3)
            v = v.permute(0, 2, 1, 3)
            x = flash_attn_func(
                q,
                k,
                v,
                dropout_p=self.attn_drop.p if self.training else 0.0,
                softmax_scale=self.scale,
            )
        else:
            dtype = q.dtype
            q = q * self.scale
            attn = q @ k.transpose(-2, -1)  # translate attn to float32
            attn = attn.to(torch.float32)
            attn = attn.softmax(dim=-1)
            attn = attn.to(dtype)  # cast back attn to original dtype
            attn = self.attn_drop(attn)
            x = attn @ v

        x_output_shape = (B, N, C)
        if not enable_flash_attn:
            x = x.transpose(1, 2)
        x = x.reshape(x_output_shape)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x


class MultiHeadCrossAttention(nn.Module):
    def __init__(self, d_model, num_heads, attn_drop=0.0, proj_drop=0.0):
        super(MultiHeadCrossAttention, self).__init__()
        assert d_model % num_heads == 0, "d_model must be divisible by num_heads"

        self.d_model = d_model
        self.num_heads = num_heads
        self.head_dim = d_model // num_heads

        self.q_linear = nn.Linear(d_model, d_model)
        self.kv_linear = nn.Linear(d_model, d_model * 2)
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(d_model, d_model)
        self.proj_drop = nn.Dropout(proj_drop)

    def forward(self, x, cond, mask=None):
        # query/value: img tokens; key: condition; mask: if padding tokens
        B, N, C = x.shape

        q = self.q_linear(x).view(1, -1, self.num_heads, self.head_dim)
        kv = self.kv_linear(cond).view(1, -1, 2, self.num_heads, self.head_dim)
        k, v = kv.unbind(2)

        attn_bias = None
        # TODO: support torch computation
        import xformers.ops

        if mask is not None:
            attn_bias = xformers.ops.fmha.BlockDiagonalMask.from_seqlens([N] * B, mask)
        x = xformers.ops.memory_efficient_attention(q, k, v, p=self.attn_drop.p, attn_bias=attn_bias)

        x = x.view(B, -1, C)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x