import torch
import torch.nn as nn
from torch.nn import TransformerEncoder, TransformerEncoderLayer

class EvoTransformer(nn.Module):
    def __init__(self, vocab_size=30522, d_model=512, nhead=8, num_layers=6, dim_feedforward=1024, dropout=0.1):
        super(EvoTransformer, self).__init__()
        self.embedding = nn.Embedding(vocab_size, d_model)
        self.memory_token = nn.Parameter(torch.zeros(1, 1, d_model))
        
        encoder_layer = TransformerEncoderLayer(d_model=d_model, nhead=nhead, dim_feedforward=dim_feedforward, dropout=dropout)
        self.transformer = TransformerEncoder(encoder_layer, num_layers=num_layers)
        
        self.memory_proj = nn.Linear(d_model, d_model)
        self.norm = nn.LayerNorm(d_model)
        
        self.classifier = nn.Linear(d_model, 1)  # Matches saved model: output is a single logit

    def forward(self, input_ids):
        x = self.embedding(input_ids)

        memory_token = self.memory_token.expand(x.size(0), -1, -1)
        x = torch.cat([memory_token, x], dim=1)

        x = self.transformer(x)
        x = self.norm(x)

        memory_output = self.memory_proj(x[:, 0])
        return memory_output