Update modeling_InternLM_XComposer.py

model.to_empty() on CPU will help the model to successfully be loaded on 3090/4090.

modeling_InternLM_XComposer.py

@@ -74,7 +74,8 @@ class InternLMXComposerForCausalLM(PreTrainedModel):
             # speed up init llm
             with torch.device('meta'):
                 self.internlm_model = InternLMForCausalLM._from_config(config)
-            self.internlm_model.to_empty(device=config.device).to(torch.float16)
+            self.internlm_model.to_empty(device='cpu').to(torch.float16)
+            self.internlm_model.to(config.device)
         for n, m in self.internlm_model.named_modules():
             if 'lora' in n:
                 m.float()