update flask

gradio_demo/seed_llama_flask.py

@@ -66,8 +66,12 @@ class LLMService:
         self.image_id_shift = 32000
 
         self.image_transform = hydra.utils.instantiate(image_transform_cfg)
-        self.tokenizer = hydra.utils.instantiate(tokenizer_cfg, device=args.tokenizer_device, load_diffusion=True)
 
+        model = hydra.utils.instantiate(model_cfg, device_map=args.llm_device).eval()
+        self.model = model
+        print(model.get_memory_footprint())
+
+        self.tokenizer = hydra.utils.instantiate(tokenizer_cfg, device=args.tokenizer_device, load_diffusion=True)
         if args.offload_encoder:
             self.tokenizer.image_tokenizer.model.visual_encoder.to('cpu')
         if args.offload_decoder:
@@ -75,9 +79,6 @@ class LLMService:
             
         # model = hydra.utils.instantiate(model_cfg, torch_dtype=torch.float16)
         # self.model = model.eval().to(args.llm_device)
-        model = hydra.utils.instantiate(model_cfg, device_map=args.llm_device).eval()
-        self.model = model
-        print(model.get_memory_footprint())
         self.llm_device = args.llm_device
         self.tokenizer_device = args.tokenizer_device
         self.offload_encoder = args.offload_encoder