Update app.py

app.py

@@ -40,15 +40,15 @@ tokenizer = AutoTokenizer.from_pretrained(
     , token= token,)
 
 
-# with init_empty_weights():
-model = AutoModelForCausalLM.from_pretrained(model_id, token= token, 
+with init_empty_weights():
+    model = AutoModelForCausalLM.from_pretrained(model_id, token= token, 
                                                  # torch_dtype= torch.uint8, 
                                              torch_dtype=torch.float16,
                                               # torch_dtype=torch.fl,
                                              attn_implementation="flash_attention_2",
                                              low_cpu_mem_usage=True,
                                              # llm_int8_enable_fp32_cpu_offload=True,
-                                             # device_map='cuda',
+                                             device_map='cuda',
                                              
                                             )
 
@@ -56,7 +56,7 @@ model = AutoModelForCausalLM.from_pretrained(model_id, token= token,
 # 
 
 
-# device_map = infer_auto_device_map(model, max_memory={0: "80GB", 1: "80GB", "cpu": "65GB"})
+device_map = infer_auto_device_map(model, max_memory={0: "79GB", })
 
 # 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"])