Adding paper

models/blocks.py

@@ -125,6 +125,7 @@ class TransformerBlock(nn.Module):
         self,
         x: torch.Tensor,
         mask: Optional[torch.Tensor] = None,
+        use_cache: bool = False,
     ) -> torch.Tensor:
         """
         Forward pass through transformer block.
@@ -132,6 +133,7 @@ class TransformerBlock(nn.Module):
         Args:
             x: Input tensor [batch_size, seq_len, d_model]
             mask: Optional attention mask
+            use_cache: Whether to use KV cache (for optimized attention)
             
         Returns:
             Output tensor [batch_size, seq_len, d_model]
@@ -139,7 +141,11 @@ class TransformerBlock(nn.Module):
         # Pre-norm self-attention with residual connection
         residual = x
         x = self.norm1(x)
-        attn_out, _ = self.self_attn(x, x, x, mask=mask)
+        # Pass use_cache to attention layer if it's OptimizedMultiHeadAttention
+        if isinstance(self.self_attn, OptimizedMultiHeadAttention):
+            attn_out, _ = self.self_attn(x, x, x, mask=mask, use_cache=use_cache)
+        else:
+            attn_out, _ = self.self_attn(x, x, x, mask=mask)
         x = residual + self.dropout(attn_out)
         
         # Pre-norm feed-forward with residual connection

models/optimized_attention.py

@@ -119,8 +119,17 @@ class OptimizedMultiHeadAttention(nn.Module):
         
         # Use KV cache if available and enabled
         if use_cache and self.kv_cache is not None:
-            # Append new keys and values to cache
+            # Get current cache length
-            self.kv_cache.append(K, V)
+            cached_len = self.kv_cache.keys.shape[2] if self.kv_cache.keys.numel() > 0 else 0
+            
+            # Only append new tokens (those after the cached length)
+            if cached_len < seq_len:
+                # Extract only the new tokens to append
+                new_k = K[:, :, cached_len:, :]  # [batch_size, num_heads, new_seq_len, d_k]
+                new_v = V[:, :, cached_len:, :]
+                self.kv_cache.append(new_k, new_v)
+            
+            # Use all cached keys and values
             K = self.kv_cache.keys
             V = self.kv_cache.values
             kv_seq_len = K.shape[2]
@@ -328,9 +337,11 @@ class OptimizedInference:
         
         generated = input_ids.clone()
         
+        # Process initial prompt to populate cache
+        logits, _ = self.model(generated, use_cache=True)
+        
         for _ in range(max_length - input_ids.shape[1]):
-            # Forward pass
+            # Get next token logits from last position
-            logits, _ = self.model(generated)
             next_token_logits = logits[:, -1, :] / temperature
             
             # Apply top-k filtering
@@ -370,6 +381,9 @@ class OptimizedInference:
             
             # Append to generated sequence
             generated = torch.cat([generated, next_token], dim=1)
+            
+            # Forward pass with cache enabled (processes full sequence but uses cached K/V)
+            logits, _ = self.model(generated, use_cache=True)
         
         # Clear KV cache
         for module in self.model.modules():

models/transformer.py

@@ -129,6 +129,7 @@ class TransformerModel(nn.Module):
         self,
         input_ids: torch.Tensor,
         attention_mask: Optional[torch.Tensor] = None,
+        use_cache: bool = False,
     ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
         """
         Forward pass through the transformer model.
@@ -136,6 +137,7 @@ class TransformerModel(nn.Module):
         Args:
             input_ids: Token indices [batch_size, seq_len]
             attention_mask: Optional attention mask [batch_size, seq_len]
+            use_cache: Whether to use KV cache (for optimized attention)
             
         Returns:
             logits: Output logits [batch_size, seq_len, vocab_size]
@@ -166,7 +168,7 @@ class TransformerModel(nn.Module):
         
         # Pass through transformer blocks
         for layer in self.layers:
-            x = layer(x, mask=attention_mask)
+            x = layer(x, mask=attention_mask, use_cache=use_cache)
         
         # Final layer norm
         x = self.final_norm(x)