Add the option to tie_word_embeddings (#1260)

Branch: GraniteCodeSupport

Signed-off-by: Gabe Goodhart <[email protected]>

Author: gabe-l-hart

torchchat/model.py

@@ -281,6 +281,8 @@ class TransformerArgs:
     # Optional biases
     attention_bias: bool = False
     feed_forward_bias: bool = False
+    # Whether or not to tie the input word embeddings to the output
+    tie_word_embeddings: bool = False
 
     def __post_init__(self):
         if self.n_local_heads == -1:
@@ -632,12 +634,20 @@ def __init__(self, config: TransformerArgs) -> None:
         if config.stage_idx == config.n_stages - 1:
             self.norm = RMSNorm(config.dim, eps=config.norm_eps)
             self.output = nn.Linear(config.dim, config.vocab_size, bias=False)
+            if config.tie_word_embeddings:
+                self.output.weight = self.tok_embeddings.weight
         else:
             self.norm = None
             self.output = None
 
         self.max_batch_size = -1
         self.max_seq_length = -1
+        self._register_load_state_dict_pre_hook(self.load_hook)
+
+    def load_hook(self, state_dict, prefix, *args):
+        """Handle tied embeddings at load time"""
+        if self.config.tie_word_embeddings:
+            state_dict.setdefault("model.output.weight", state_dict["model.tok_embeddings.weight"])
 
     def setup_caches(self, max_batch_size, max_seq_length, cache_lanes: int = 1):
         if (