Update app.py

app.py

@@ -5,78 +5,38 @@ from transformers import AutoTokenizer, AutoModelForCausalLM
 # Configurações do modelo DeepSeek-R1
 MODEL_NAME = "deepseek-ai/DeepSeek-R1"
 
-# Carregar tokenizer e modelo; Configurar para CPU
+# Configuração segura para CPU
 tokenizer = AutoTokenizer.from_pretrained(
     MODEL_NAME,
-    trust_remote_code=True
+    trust_remote_code=True,
+    revision="6528ae3"  # Fixar versão específica
 )
 
 model = AutoModelForCausalLM.from_pretrained(
     MODEL_NAME,
     trust_remote_code=True,
     torch_dtype=torch.float32,
-    device_map="auto",
-    low_cpu_mem_usage=True
+    device_map="cpu",
+    low_cpu_mem_usage=True,
+    load_in_8bit=False,
+    offload_folder="offload"  # Pasta para descarregar pesos grandes
 )
 
-# Configurações de comprimento
-TARGET_LENGTH = 256
-MARGIN = 6
-MIN_LENGTH = TARGET_LENGTH - MARGIN
-MAX_LENGTH = TARGET_LENGTH + MARGIN
-MAX_ATTEMPTS = 5
-
 def summarize_text(text):
-    """
-    Gera resumo adaptado para o DeepSeek-R1 com ajuste de comprimento
-    """
-    best_summary = ""
-    best_distance = float("inf")
-    adjusted_max_tokens = 512  # Valor inicial ajustável
-
-    for attempt in range(MAX_ATTEMPTS):
-        # Formatar prompt para sumarização
-        prompt = f"Resuma o seguinte texto em português com cerca de {TARGET_LENGTH} caracteres:\n{text}\nResumo:"
-        
-        inputs = tokenizer.encode(
-            prompt,
-            return_tensors="pt",
-            max_length=4096,  # Ajustar conforme capacidade do modelo
-            truncation=True
+    prompt = f"Resuma em português ({TARGET_LENGTH} caracteres): {text}"
+    inputs = tokenizer(prompt, return_tensors="pt", max_length=2048, truncation=True)
+    
+    with torch.inference_mode():
+        outputs = model.generate(
+            **inputs,
+            max_new_tokens=MAX_LENGTH,
+            temperature=0.9,
+            top_k=50,
+            no_repeat_ngram_size=3
         )
-        
-        # Gerar sumário
-        summary_ids = model.generate(
-            inputs,
-            max_new_tokens=adjusted_max_tokens,
-            num_beams=5,
-            repetition_penalty=1.2,
-            early_stopping=True,
-            temperature=0.7,
-            top_p=0.9
-        )
-        
-        summary = tokenizer.decode(summary_ids[0], skip_special_tokens=True)
-        
-        # Filtrar apenas o resumo gerado (remover prompt)
-        if "Resumo:" in summary:
-            summary = summary.split("Resumo:")[-1].strip()
-        
-        summary_length = len(summary)
-        distance = abs(TARGET_LENGTH - summary_length)
-        
-        if distance < best_distance:
-            best_summary = summary
-            best_distance = distance
-            
-        if MIN_LENGTH <= summary_length <= MAX_LENGTH:
-            return summary[:MAX_LENGTH]  # Garantir limite máximo
-        
-        # Ajuste adaptativo
-        adjustment = int((summary_length / TARGET_LENGTH) * adjusted_max_tokens)
-        adjusted_max_tokens = max(32, adjusted_max_tokens - adjustment)
-
-    return best_summary[:MAX_LENGTH]
+    
+    summary = tokenizer.decode(outputs[0], skip_special_tokens=True)
+    return summary[len(prompt):].strip()
 
 # Interface Gradio
 interface = gr.Interface(