perf: use safetensors for faster model loading.

app.py

@@ -73,14 +73,16 @@ def load_model_on_gpu():
             device_map="auto",
             torch_dtype=torch.float16,
             trust_remote_code=True,
-            low_cpu_mem_usage=True
+            low_cpu_mem_usage=True,
+            use_safetensors=True
         )
         
         # Load LoRA adapter
         model = PeftModel.from_pretrained(
             base_model, 
             LORA_MODEL,
-            torch_dtype=torch.float16
+            torch_dtype=torch.float16,
+            use_safetensors=True
         )
         model.eval()