Fix qwen accuracy issue (#2945)

python/sglang/srt/models/qwen2.py

@@ -356,8 +356,6 @@ class Qwen2ForCausalLM(nn.Module):
                 break
             else:
                 # Skip loading extra bias for GPTQ models.
-                if "lm_head.weight" in name:
-                    continue
                 if name.endswith(".bias") and name not in params_dict:
                     continue
                 param = params_dict[name]