[Bugfix] Fix the lm_head in gpt_bigcode in lora mode (#6357)

Signed-off-by: Max de Bayser <mbayser@br.ibm.com>
Signed-off-by: Max de Bayser <maxdebayser@gmail.com>

vllm/model_executor/models/gpt_bigcode.py

@@ -272,12 +272,6 @@ class GPTBigCodeModel(nn.Module):
 class GPTBigCodeForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
     packed_modules_mapping = {"c_attn": ["c_attn"]}
 
-    # LoRA specific attributes
-    embedding_modules = {
-        "wte": "input_embeddings",
-        "lm_head": "output_embeddings",
-    }
-
     def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         super().__init__()
         config = vllm_config.model_config.hf_config
@@ -330,8 +324,11 @@ class GPTBigCodeForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
 
     def load_weights(self, weights: Iterable[tuple[str,
                                                    torch.Tensor]]) -> set[str]:
+        skip_prefixes = None
+        if self.config.tie_word_embeddings:
+            skip_prefixes = ["lm_head."]
         loader = AutoWeightsLoader(
             self,
-            skip_prefixes=(["lm_head."]),
+            skip_prefixes=skip_prefixes,
         )
-        return loader.load_weights(weights)
\ No newline at end of file
+        return loader.load_weights(weights)