to_empty('meta') --> to_empty('cuda')

modeling_InternLM_XComposer.py

@@ -67,7 +67,7 @@ class InternLMXComposerForCausalLM(PreTrainedModel):
             # speed up init llm
             with torch.device('meta'):
                 self.internlm_model = InternLMForCausalLM._from_config(config)
-            self.internlm_model.to_empty(device='meta').to(torch.float16)
+            self.internlm_model.to_empty(device=config.device).to(torch.float16)
             self.internlm_model.tie_weights()
             # self.internlm_model.to(config.device)