Update evo_model.py

evo_model.py

@@ -1,41 +1,47 @@
-# evo_model.py — Defines EvoDecoderModel used in inference and training
+# evo_model.py — EvoDecoderModel with fixed positional encoding (max_len=512)
 import torch
 import torch.nn as nn
-import math
 
+# Positional encoding used by transformer decoders
 class PositionalEncoding(nn.Module):
-    def __init__(self, d_model, max_len=128):
+    def __init__(self, d_model, max_len=512):  # Increased max_len
         super().__init__()
         pe = torch.zeros(max_len, d_model)
-        position = torch.arange(0, max_len).unsqueeze(1)
-        div_term = torch.exp(torch.arange(0, d_model, 2) * (-math.log(10000.0) / d_model))
+        position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
+        div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-torch.log(torch.tensor(10000.0)) / d_model))
         pe[:, 0::2] = torch.sin(position * div_term)
         pe[:, 1::2] = torch.cos(position * div_term)
-        pe = pe.unsqueeze(0)  # [1, max_len, d_model]
+        pe = pe.unsqueeze(0)  # shape: (1, max_len, d_model)
         self.register_buffer('pe', pe)
 
     def forward(self, x):
-        return x + self.pe[:, :x.size(1)]
+        return x + self.pe[:, :x.size(1)]  # Match sequence length
 
+# Main EvoDecoder model
 class EvoDecoderModel(nn.Module):
-    def __init__(self, vocab_size, d_model=512, nhead=8, num_layers=6, dim_ff=2048, max_len=128):
+    def __init__(self, vocab_size, d_model=384, nhead=6, num_layers=6, dim_feedforward=1024, dropout=0.1):
         super().__init__()
-        self.token_embed = nn.Embedding(vocab_size, d_model)
-        self.pos_encoder = PositionalEncoding(d_model, max_len)
-
-        decoder_layer = nn.TransformerDecoderLayer(d_model, nhead, dim_ff, batch_first=True)
+        self.embedding = nn.Embedding(vocab_size, d_model)
+        self.pos_encoder = PositionalEncoding(d_model)
+        decoder_layer = nn.TransformerDecoderLayer(
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            dropout=dropout,
+            batch_first=True
+        )
         self.decoder = nn.TransformerDecoder(decoder_layer, num_layers=num_layers)
-
-        self.lm_head = nn.Linear(d_model, vocab_size)
-
-    def generate_square_subsequent_mask(self, sz):
-        return torch.triu(torch.full((sz, sz), float('-inf')), diagonal=1)
+        self.linear = nn.Linear(d_model, vocab_size)
 
     def forward(self, input_ids):
-        x = self.token_embed(input_ids)
-        x = self.pos_encoder(x)
+        embedded = self.embedding(input_ids)
+        embedded = self.pos_encoder(embedded)
 
-        tgt_mask = self.generate_square_subsequent_mask(x.size(1)).to(x.device)
-        x = self.decoder(x, x, tgt_mask=tgt_mask)
+        # Create causal mask for autoregressive decoding
+        seq_len = embedded.size(1)
+        mask = torch.triu(torch.ones(seq_len, seq_len, device=embedded.device), diagonal=1).bool()
 
-        return self.lm_head(x)
+        # Use the input itself as memory for self-decoding
+        output = self.decoder(embedded, embedded, tgt_mask=mask)
+        logits = self.linear(output)
+        return logits