expand coverage of gpt2 model loading (#271)

vllm/model_executor/models/gpt2.py

@@ -228,10 +228,12 @@ class GPT2LMHeadModel(nn.Module):
                 # GPT-2 ties the weights of the embedding layer and the final
                 # linear layer.
                 continue
-            if ".attn.bias" in name:
+            if ".attn.bias" in name or ".attn.masked_bias" in name:
                 # Skip attention mask.
                 # NOTE: "c_attn.bias" should not be skipped.
                 continue
+
+            if not name.startswith("transformer."):
                 name = "transformer." + name
 
             # The HF's GPT-2 implementation uses Conv1D instead of Linear.