Update app.py

app.py

@@ -39,6 +39,8 @@ tokenizer = AutoTokenizer.from_pretrained(
     model_id
     , token= token,)
 
+device_map = infer_auto_device_map(model, max_memory={0: "80GB", "cpu": "65GB"})
+
 with init_empty_weights():
     model = AutoModelForCausalLM.from_pretrained(model_id, token= token, 
                                                  # torch_dtype= torch.uint8, 
@@ -47,7 +49,7 @@ with init_empty_weights():
                                              attn_implementation="flash_attention_2",
                                              # low_cpu_mem_usage=True,
                                              # llm_int8_enable_fp32_cpu_offload=True,
-                                             # device_map="cuda",
+                                             device_map=device_map,
                                              
                                             )
 
@@ -58,8 +60,8 @@ with init_empty_weights():
 device_map = infer_auto_device_map(model, max_memory={0: "80GB", 1: "80GB", "cpu": "65GB"})
 
 # Load the model with the inferred device map
-model = load_checkpoint_and_dispatch(model, model_id, device_map=device_map, no_split_module_classes=["GPTJBlock"])
-model.half()
+# model = load_checkpoint_and_dispatch(model, model_id, device_map=device_map, no_split_module_classes=["GPTJBlock"])
+# model.half()