bugfix: llava-hf/llava-interleave-qwen-7b-hf (InternLM#2657)

- fix init raise exception because tie_word_embeddings config

lmdeploy/vl/model/llava_hf.py

@@ -31,13 +31,15 @@ def build_model(self):
             else:
                 self.vl_model = model
 
+        # fix for llava-hf/llava-interleave-qwen-7b-hf
+        setattr(model.config, "tie_word_embeddings", False)
         with disable_logging():
             load_checkpoint_and_dispatch(
                 model=model,
                 max_memory=self.max_memory,
                 checkpoint=self.model_path,
                 device_map='auto' if not self.with_llm else {'': 'cpu'},
-                no_split_module_classes=['CLIPEncoderLayer'],
+                no_split_module_classes=['CLIPEncoderLayer', 'SiglipEncoderLayer'],
                 dtype=torch.half)
         model.eval()
         self.model = model