feat: update llm-webkit extract

Author: e06084

webmainbench/extractors/llm_webkit_extractor.py

@@ -370,20 +370,16 @@ def _load_vllm_model(self):
                 trust_remote_code=True
             )
 
-            # vLLM配置
+            # vLLM配置 - 参考ray_test_qa.py的简化配置
             model_kwargs = {
                 "model": self.inference_config.model_path,
                 "trust_remote_code": True,
                 "dtype": self.inference_config.dtype,
                 "tensor_parallel_size": self.inference_config.tensor_parallel_size,
-                "max_model_len": self.inference_config.max_tokens,
-                "max_num_batched_tokens": max(self.inference_config.max_tokens, 8192),
-                "gpu_memory_utilization": self.inference_config.gpu_memory_utilization,
-                "enforce_eager": self.inference_config.enforce_eager,
-                "disable_custom_all_reduce": True,
-                "load_format": "auto",
             }
 
+            print(f"🔧 vLLM配置: {model_kwargs}")
+            
             self.model = LLM(**model_kwargs)
 
             # 初始化token状态管理器
@@ -397,8 +393,8 @@ def _load_vllm_model(self):
             print("✅ vLLM模型加载成功!")
 
         except Exception as e:
-            print(f"⚠️  vLLM加载失败，回退到transformers: {e}")
-            self._load_transformers_model()
+            print(f"❌ vLLM加载失败: {e}")
+            raise RuntimeError(f"vLLM模型加载失败: {e}")
 
     def _create_prompt(self, simplified_html: str) -> str:
         """创建分类提示."""