增加昇腾去初始化操作

example/benchmark/benchmark.cpp

@@ -198,5 +198,6 @@ int main(int argc, char **argv) {
     printf("prompt use %f s\n", promptSpend);
     printf("prompt speed = %f tokens / s\n", (float)promptTokenNum / promptSpend);
     printf("output %d tokens\nuse %f s\nspeed = %f tokens / s\n", tokens, spend, tokens / spend);
+    fastllm::Finalize();
     return 0;
 }

include/fastllm.h

@@ -508,6 +508,8 @@ namespace fastllm {
 
     void PrintProfiler();
 
+    void Finalize(); // 程序执行完，释放资源
+
     void ApplyDeviceMap(const std::map <std::string, int> &deviceMap, int current, int total); // 执行到了current, 一共total，使用deviceMap切换设备
 
     int LLMSamplingOnly(Data &logits, int outerOffset, const GenerationConfig &config);

main.cpp

@@ -149,6 +149,6 @@ int main(int argc, char **argv) {
         }, generationConfig);
         messages.push_back(std::make_pair("assistant", ret));
     }
-
+    fastllm::Finalize();
     return 0;
 }

src/devices/ascend/fastllm-acl.cpp

@@ -21,8 +21,10 @@ void showError(aclError state, char const* const message, const char* const file
            int const line) {
     if (ACL_SUCCESS != state) {
         printf("%s  AsecndCL error = %d at %s:%d\n", message, state, file, line);
-        if (state >= 500000)
+        if (state >= 500000) {
+            aclFinalize();
             exit(state);
+        }
     }  
 }
 

src/fastllm.cpp

@@ -3140,4 +3140,10 @@ namespace fastllm {
     std::map <std::string, int> GetDeviceMap() {
         return defaultDeviceMap;
     }
+
+    void Finalize() {
+#ifdef USE_ASCEND_NPU
+        npu::FastllmAclFinalize();
+#endif
+    }
 }