evo_decoder.py

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

class SelfAttention(nn.Module):
    def __init__(self, d_model, nhead):
        super().__init__()
        self.qkv_proj = nn.Linear(d_model, 3 * d_model)
        self.out_proj = nn.Linear(d_model, d_model)
        self.nhead = nhead
        self.d_model = d_model

    def forward(self, x):
        B, T, C = x.size()
        qkv = self.qkv_proj(x)
        q, k, v = qkv.chunk(3, dim=-1)

        q = q.view(B, T, self.nhead, C // self.nhead).transpose(1, 2)
        k = k.view(B, T, self.nhead, C // self.nhead).transpose(1, 2)
        v = v.view(B, T, self.nhead, C // self.nhead).transpose(1, 2)

        scores = torch.matmul(q, k.transpose(-2, -1)) / (C // self.nhead) ** 0.5
        attn = torch.softmax(scores, dim=-1)
        out = torch.matmul(attn, v)

        out = out.transpose(1, 2).contiguous().view(B, T, C)
        return self.out_proj(out)

class FeedForward(nn.Module):
    def __init__(self, d_model, dim_feedforward):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(d_model, dim_feedforward),
            nn.ReLU(),
            nn.Dropout(),  # ✅ Important: was present in the training model
            nn.Linear(dim_feedforward, d_model)
        )

    def forward(self, x):
        return self.net(x)

class TransformerBlock(nn.Module):
    def __init__(self, d_model, nhead, dim_feedforward):
        super().__init__()
        self.attn = SelfAttention(d_model, nhead)
        self.ln1 = nn.LayerNorm(d_model)
        self.ffn = FeedForward(d_model, dim_feedforward)
        self.ln2 = nn.LayerNorm(d_model)

    def forward(self, x):
        x = x + self.attn(self.ln1(x))
        x = x + self.ffn(self.ln2(x))
        return x

class EvoDecoder(nn.Module):
    def __init__(self, vocab_size, d_model=256, nhead=4, num_layers=3, dim_feedforward=512):
        super().__init__()
        self.token_emb = nn.Embedding(vocab_size, d_model)
        self.pos_emb = nn.Embedding(512, d_model)
        self.blocks = nn.Sequential(*[
            TransformerBlock(d_model, nhead, dim_feedforward) for _ in range(num_layers)
        ])
        self.ln_f = nn.LayerNorm(d_model)
        self.fc_out = nn.Linear(d_model, vocab_size)

    def forward(self, x):
        B, T = x.size()
        tok = self.token_emb(x)
        pos = self.pos_emb(torch.arange(T, device=x.device).unsqueeze(0).expand(B, T))
        x = tok + pos
        x = self.blocks(x)
        x = self.ln_f(x)
        return self.fc_out(x)