# Adapted from https://github.com/louaaron/Score-Entropy-Discrete-Diffusion/blob/main/model/transformer.py
# Not used in this repo but implemented for future reference using flash attention


import torch
import torch.nn as nn
import torch.nn.functional as F
import math

from einops import rearrange
from flash_attn.flash_attn_interface import flash_attn_varlen_qkvpacked_func
from huggingface_hub import PyTorchModelHubMixin

from . import rotary
from .fused_add_dropout_scale import (
    bias_dropout_add_scale_fused_train, 
    bias_dropout_add_scale_fused_inference, 
    modulate_fused,
)

from .dit import LayerNorm, TimestepEmbedder


#################################################################################
#                                 Core Model                                    #
#################################################################################


class DiTBlock(nn.Module):

    def __init__(self, dim, n_heads, cond_dim, mlp_ratio=4, dropout=0.1):
        super().__init__()
        self.n_heads = n_heads

        self.norm1 = LayerNorm(dim)
        self.attn_qkv = nn.Linear(dim, 3 * dim, bias=False)
        self.attn_out = nn.Linear(dim, dim, bias=False)
        self.dropout1 = nn.Dropout(dropout)

        self.norm2 = LayerNorm(dim)
        self.mlp = nn.Sequential(
            nn.Linear(dim, mlp_ratio * dim, bias=True),
            nn.GELU(approximate="tanh"),
            nn.Linear(mlp_ratio * dim, dim, bias=True)
        )
        self.dropout2 = nn.Dropout(dropout)

        self.dropout = dropout
        

        self.adaLN_modulation = nn.Linear(cond_dim, 6 * dim, bias=True)
        self.adaLN_modulation.weight.data.zero_()
        self.adaLN_modulation.bias.data.zero_()


    def _get_bias_dropout_scale(self):
        return (
            bias_dropout_add_scale_fused_train
            if self.training
            else bias_dropout_add_scale_fused_inference
        )


    def forward(self, x, rotary_cos_sin, c, seqlens=None):
        batch_size, seq_len = x.shape[0], x.shape[1]

        bias_dropout_scale_fn = self._get_bias_dropout_scale()

        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(c)[:, None].chunk(6, dim=2)

        # attention operation
        x_skip = x
        x = modulate_fused(self.norm1(x), shift_msa, scale_msa)
        # dtype0 = x.dtype

        qkv = self.attn_qkv(x)
        qkv = rearrange(qkv, 'b s (three h d) -> b s three h d', three=3, h=self.n_heads)
        with torch.cuda.amp.autocast(enabled=False):
            cos, sin = rotary_cos_sin
            qkv = rotary.apply_rotary_pos_emb_flash(
                qkv, cos.to(qkv.dtype), sin.to(qkv.dtype)
            )
        qkv = rearrange(qkv, 'b s ... -> (b s) ...')
        if seqlens is None:
            cu_seqlens = torch.arange(
                0, (batch_size + 1) * seq_len, step=seq_len,
                dtype=torch.int32, device=qkv.device
            )
        else:
            cu_seqlens = seqlens.cumsum(-1)

        x = flash_attn_varlen_qkvpacked_func(
            qkv, cu_seqlens, seq_len, 0., causal=False)
        
        x = rearrange(x, '(b s) h d -> b s (h d)', b=batch_size)

        x = bias_dropout_scale_fn(self.attn_out(x), None, gate_msa, x_skip, self.dropout)

        # mlp operation
        x = bias_dropout_scale_fn(self.mlp(modulate_fused(self.norm2(x), shift_mlp, scale_mlp)), None, gate_mlp, x, self.dropout)
        return x



class EmbeddingLayer(nn.Module):
    def __init__(self, dim, vocab_dim):
        """
        Mode arg: 0 -> use a learned layer, 1 -> use eigenvectors, 
        2-> add in eigenvectors, 3 -> use pretrained embedding matrix
        """
        super().__init__()
        # self.embedding = nn.Parameter(torch.empty((vocab_dim, dim)))
        self.embedding = nn.Linear(vocab_dim, dim)
        torch.nn.init.kaiming_uniform_(self.embedding.weight, a=math.sqrt(5))

    def forward(self, x):
        # return self.embedding[x]
        return self.embedding(x)


class DiTFinalLayer(nn.Module):
    def __init__(self, hidden_size, out_channels, cond_dim):
        super().__init__()
        self.norm_final = LayerNorm(hidden_size)
        self.linear = nn.Linear(hidden_size, out_channels)
        self.linear.weight.data.zero_()
        self.linear.bias.data.zero_()

        self.adaLN_modulation = nn.Linear(cond_dim, 2 * hidden_size, bias=True)
        self.adaLN_modulation.weight.data.zero_()
        self.adaLN_modulation.bias.data.zero_()


    def forward(self, x, c):
        shift, scale = self.adaLN_modulation(c)[:, None].chunk(2, dim=2)
        x = modulate_fused(self.norm_final(x), shift, scale)
        x = self.linear(x)
        return x


class DiT(nn.Module, PyTorchModelHubMixin):
    def __init__(
            self,
            input_dim: int,
            output_dim: int,
            hidden_size: int,
            n_heads: int,
            cond_dim: int,
            dropout: float,
            n_blocks: int,
            **kwargs
    ):
        super().__init__()

        self.input_dim = input_dim
        self.output_dim = output_dim

        self.vocab_embed = EmbeddingLayer(hidden_size, input_dim)
        self.sigma_map = TimestepEmbedder(cond_dim)
        self.rotary_emb = rotary.Rotary(hidden_size // n_heads)

        self.blocks = nn.ModuleList([
            DiTBlock(hidden_size, n_heads, cond_dim, dropout=dropout) 
            for _ in range(n_blocks)
        ])

        self.output_layer = DiTFinalLayer(hidden_size, output_dim, cond_dim)
        self.kwargs = kwargs

    def _get_bias_dropout_scale(self):
        return (
            bias_dropout_add_scale_fused_train
            if self.training
            else bias_dropout_add_scale_fused_inference
        )

    def forward(self, input_vector, t):
        x = self.vocab_embed(input_vector)
        c = F.silu(self.sigma_map(t))

        rotary_cos_sin = self.rotary_emb(x)

        with torch.cuda.amp.autocast(dtype=torch.bfloat16):
            for i in range(len(self.blocks)):
                x = self.blocks[i](x, rotary_cos_sin, c, seqlens=None)
            x = self.output_layer(x, c) # BxCxD

        return x