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

class EvoEncoder(nn.Module):
    def __init__(self, d_model=512, num_heads=8, ffn_dim=1024, num_layers=6, memory_enabled=True):
        super().__init__()
        self.embedding = nn.Embedding(30522, d_model)
        self.memory_enabled = memory_enabled
        if memory_enabled:
            self.memory_proj = nn.Linear(d_model, d_model)
            self.memory_token = nn.Parameter(torch.zeros(1, 1, d_model))
        else:
            self.memory_token = None

        encoder_layer = nn.TransformerEncoderLayer(
            d_model=d_model,
            nhead=num_heads,
            dim_feedforward=ffn_dim,
            batch_first=True
        )
        self.transformer = nn.TransformerEncoder(encoder_layer, num_layers=num_layers)

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

        if self.memory_enabled and self.memory_token is not None:
            mem = self.memory_token.expand(x.size(0), 1, x.size(2))
            x = torch.cat([mem, x], dim=1)

        x = self.transformer(x)
        return x

class EvoTransformerV22(nn.Module):
    def __init__(self):
        super().__init__()
        self.encoder = EvoEncoder(d_model=512, num_heads=8, ffn_dim=1024, num_layers=6, memory_enabled=True)
        self.pool = nn.AdaptiveAvgPool1d(1)
        self.classifier = nn.Linear(512, 1)  # ✅ Matches checkpoint

    def forward(self, input_ids):
        x = self.encoder(input_ids)
        x = self.pool(x.transpose(1, 2)).squeeze(-1)
        return self.classifier(x)  # Output: [batch_size, 1]