fix zero

model_lib/modules.py

@@ -84,11 +84,11 @@ class MoMA_main_modal(nn.Module):
         
         print('Loading MoMA: its Multi-modal LLM...')
         model_name = get_model_name_from_path(args.model_path)
-        self.tokenizer_llava, self.model_llava, self.image_processor_llava, self.context_len_llava = load_pretrained_model(args.model_path, None, model_name, load_8bit=self.args.load_8bit, load_4bit=self.args.load_4bit, device=args.device)
+        self.tokenizer_llava, self.model_llava, self.image_processor_llava, self.context_len_llava = load_pretrained_model(args.model_path, None, model_name, load_8bit=self.args.load_8bit, load_4bit=self.args.load_4bit)
         
         add_function(self.model_llava)
 
-        self.mapping = LlamaMLP_mapping(4096,1024).to(self.device, dtype=torch.float16)
+        self.mapping = LlamaMLP_mapping(4096,1024)
         self.load_saved_components()
         self.freeze_modules()
 

model_lib/utils.py

@@ -10,7 +10,7 @@ def parse_args():
     parser.add_argument("--model_path",type=str,default="KunpengSong/MoMA_llava_7b",help="fine tuned llava (Multi-modal LLM decoder)")
     args = parser.parse_known_args()[0]
     args.device = torch.device("cuda", 0)
-    args.load_8bit, args.load_4bit = False, True
+    args.load_8bit, args.load_4bit = False, False
     return args
 
 def show_PIL_image(tensor):