Update app.py

app.py

@@ -10,7 +10,12 @@ from langchain_community.document_loaders import PyMuPDFLoader
 # Configurações do modelo
 MODEL_PATH = "numind/NuExtract-1.5-tiny"
 tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH)
-model = AutoModelForCausalLM.from_pretrained(MODEL_PATH, torch_dtype=torch.float32)
+device = "cuda"
+model = AutoModelForCausalLM.from_pretrained(
+    MODEL_PATH, 
+    torch_dtype=torch.float16  # Usa FP16 para reduzir o uso de VRAM
+).to(device)
+torch.cuda.empty_cache()
 
 # Constantes
 MAX_INPUT_SIZE = 4000