Update app.py

app.py

@@ -1,4 +1,3 @@
-import json
 import re
 import requests
 import pandas as pd
@@ -6,6 +5,7 @@ import torch
 import gradio as gr
 from tqdm import tqdm
 from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
+import json  # Добавлен отсутствующий импорт
 
 # Конфигурация
 DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
@@ -15,40 +15,42 @@ class GAIAExpertAgent:
     def __init__(self, model_name: str = MODEL_NAME):
         self.device = "cuda" if torch.cuda.is_available() else "cpu"
         print(f"⚡ Инициализация агента на {self.device.upper()}")
+        
+        # Оптимизация загрузки модели
         self.tokenizer = AutoTokenizer.from_pretrained(model_name)
         self.model = AutoModelForSeq2SeqLM.from_pretrained(
             model_name,
             device_map="auto",
-            torch_dtype=torch.float16 if "cuda" in self.device else torch.float32
+            torch_dtype=torch.float16 if "cuda" in self.device else torch.float32,
+            low_cpu_mem_usage=True  # Снижение потребления CPU памяти
         ).eval()
         print("✅ Агент готов")
 
     def __call__(self, question: str, task_id: str = None) -> str:
         try:
-            # Специальные обработчики для GAIA
-            if "reverse" in question.lower() or "rewsna" in question:
-                return json.dumps({"final_answer": question[::-1][:100]})
-            if "how many" in question.lower() or "сколько" in question.lower():
-                numbers = re.findall(r'\d+', question)
-                result = str(sum(map(int, numbers))) if numbers else "42"
-                return json.dumps({"final_answer": result})
-            
-            # Стандартная обработка
+            # Убраны жесткие эвристики - они мешают реальным задачам GAIA
             inputs = self.tokenizer(
-                f"GAIA Question: {question}\nAnswer:",
+                f"Solve step-by-step: {question}\nFinal Answer:",
                 return_tensors="pt",
-                max_length=256,
+                max_length=512,  # Увеличен контекст
                 truncation=True
             ).to(self.device)
 
+            # Улучшена генерация
             outputs = self.model.generate(
                 **inputs,
-                max_new_tokens=50,
-                num_beams=3,
-                early_stopping=True
+                max_new_tokens=256,  # Увеличен лимит для сложных ответов
+                num_beams=5,         # Улучшено качество поиска
+                early_stopping=True,
+                repetition_penalty=2.0  # Предотвращение циклов
             )
             
             answer = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
+            
+            # Очистка памяти CUDA
+            if "cuda" in self.device:
+                torch.cuda.empty_cache()
+                
             return json.dumps({"final_answer": answer.strip()})
         
         except Exception as e:
@@ -60,8 +62,9 @@ class EvaluationRunner:
         self.api_url = api_url
         self.questions_url = f"{api_url}/questions"
         self.submit_url = f"{api_url}/submit"
+        self.session = requests.Session()  # Сессия для повторных запросов
     
-    def run_evaluation(self, agent, username: str, agent_code: str):
+    def run_evaluation(self, agent, username: str, agent_code: str, progress=tqdm):
         # Получение вопросов
         questions = self._fetch_questions()
         if not isinstance(questions, list):
@@ -70,7 +73,7 @@ class EvaluationRunner:
         # Обработка вопросов
         results = []
         answers = []
-        for q in tqdm(questions, desc="Processing"):
+        for q in progress(questions, desc="Processing GAIA tasks"):
             try:
                 json_response = agent(q["question"], q["task_id"])
                 response_obj = json.loads(json_response)
@@ -78,13 +81,13 @@ class EvaluationRunner:
                 
                 answers.append({
                     "task_id": q["task_id"],
-                    "submitted_answer": str(answer)[:300]
+                    "submitted_answer": str(answer)[:500]  # Увеличен лимит
                 })
                 
                 results.append({
                     "Task ID": q["task_id"],
-                    "Question": q["question"][:70] + "..." if len(q["question"]) > 70 else q["question"],
-                    "Answer": str(answer)[:50] + "..." if len(str(answer)) > 50 else str(answer)
+                    "Question": q["question"],
+                    "Answer": str(answer)
                 })
             except Exception as e:
                 results.append({
@@ -99,56 +102,88 @@ class EvaluationRunner:
     
     def _fetch_questions(self):
         try:
-            response = requests.get(self.questions_url, timeout=30)
+            response = self.session.get(
+                self.questions_url,
+                timeout=60,  # Увеличен таймаут
+                headers={"Accept": "application/json"}
+            )
             response.raise_for_status()
             return response.json()
         except Exception as e:
-            return f"Fetch error: {str(e)}"
+            return f"Ошибка получения вопросов: {str(e)}"
     
     def _submit_answers(self, username: str, agent_code: str, answers: list):
         try:
-            response = requests.post(
+            response = self.session.post(
                 self.submit_url,
                 json={
                     "username": username.strip(),
                     "agent_code": agent_code.strip(),
                     "answers": answers
                 },
-                timeout=60
+                timeout=120  # Увеличен таймаут
             )
             response.raise_for_status()
-            return response.json().get("message", "Answers submitted")
+            return response.json().get("message", "Ответы успешно отправлены")
         except Exception as e:
-            return f"Submission error: {str(e)}"
+            return f"Ошибка отправки: {str(e)}"
 
 
-def run_evaluation(username: str, agent_code: str):
+# Важно: Инициализация агента при запуске, а не при импорте
+def run_evaluation(username: str, agent_code: str, progress=gr.Progress()):
+    progress(0, desc="Инициализация модели...")
     agent = GAIAExpertAgent()
+    
+    progress(0, desc="Запуск оценки...")
     runner = EvaluationRunner()
-    return runner.run_evaluation(agent, username, agent_code)
+    
+    # Обертка tqdm для Gradio Progress
+    class ProgressWrapper:
+        def __call__(self, iterable, desc=""):
+            progress(0, desc=desc)
+            for i, x in enumerate(iterable):
+                progress(i / len(iterable))
+                yield x
+                
+    return runner.run_evaluation(agent, username, agent_code, progress=ProgressWrapper())
 
 
-# Интерфейс Gradio
-with gr.Blocks(title="GAIA Agent") as demo:
-    gr.Markdown("# 🧠 GAIA Agent Evaluation")
+# Оптимизированный интерфейс Gradio
+with gr.Blocks(title="GAIA Agent", theme=gr.themes.Soft()) as demo:
+    gr.Markdown("""# 🧠 GAIA Mastermind Agent
+    ## *Многошаговое решение сложных задач*""")
     
     with gr.Row():
-        with gr.Column():
+        with gr.Column(scale=1):
+            gr.Markdown("### 🔐 Авторизация")
             username = gr.Textbox(label="HF Username", value="yoshizen")
             agent_code = gr.Textbox(label="Agent Code", value="https://huggingface.co/spaces/yoshizen/FinalTest")
-            run_btn = gr.Button("Run Evaluation", variant="primary")
+            run_btn = gr.Button("Запустить оценку", variant="primary")
         
-        with gr.Column():
-            result_output = gr.Textbox(label="Status")
-            correct_output = gr.Number(label="Correct Answers")
-            total_output = gr.Number(label="Total Questions")
-            results_table = gr.Dataframe(label="Details")
+        with gr.Column(scale=2):
+            gr.Markdown("### 📊 Результаты")
+            with gr.Row():
+                result_output = gr.Textbox(label="Статус")
+                correct_output = gr.Number(label="Правильные ответы")
+                total_output = gr.Number(label="Всего вопросов")
+            results_table = gr.Dataframe(
+                label="Детализация ответов",
+                interactive=True,
+                wrap=True,
+                overflow_row_behaviour="paginate",
+                height=500
+            )
 
     run_btn.click(
         fn=run_evaluation,
         inputs=[username, agent_code],
-        outputs=[result_output, correct_output, total_output, results_table]
+        outputs=[result_output, correct_output, total_output, results_table],
+        concurrency_limit=1  # Защита от перегрузки
     )
 
 if __name__ == "__main__":
-    demo.launch(server_name="0.0.0.0", server_port=7860)
+    demo.queue(max_size=10).launch(
+        server_name="0.0.0.0",
+        server_port=7860,
+        show_error=True
+    )