fix: use engine get env

engine/cli/commands/server_start_cmd.cc

@@ -97,9 +97,9 @@ bool ServerStartCmd::Exec(const std::string& host, int port) {
       v += g;
     }
     CTL_INF("LD_LIBRARY_PATH: " << v);
-    auto data_path = file_manager_utils::GetCortexDataPath();
-    auto llamacpp_path = data_path / "engines" / "cortex.llamacpp/";
-    auto trt_path = data_path / "engines" / "cortex.tensorrt-llm/";
+    auto data_path = file_manager_utils::GetEnginesContainerPath();
+    auto llamacpp_path = data_path / "cortex.llamacpp/";
+    auto trt_path = data_path / "cortex.tensorrt-llm/";
     auto new_v = trt_path.string() + ":" + llamacpp_path.string() + ":" + v;
     setenv(name, new_v.c_str(), true);
     CTL_INF("LD_LIBRARY_PATH: " << getenv(name));

engine/services/inference_service.cc

@@ -116,6 +116,7 @@ InferResult InferenceService::LoadModel(
                ? getenv("ENGINE_PATH")
                : file_manager_utils::GetCortexDataPath().string()) +
           get_engine_path(ne);
+      LOG_INFO << "engine path: " << abs_path;
 #if defined(_WIN32)
       // TODO(?) If we only allow to load an engine at a time, the logic is simpler.
       // We would like to support running multiple engines at the same time. Therefore,

engine/utils/file_manager_utils.h

@@ -252,7 +252,9 @@ inline std::filesystem::path GetModelsContainerPath() {
 }
 
 inline std::filesystem::path GetEnginesContainerPath() {
-  auto cortex_path = GetCortexDataPath();
+  auto cortex_path = getenv("ENGINE_PATH")
+                         ? std::filesystem::path(getenv("ENGINE_PATH"))
+                         : GetCortexDataPath();
   auto engines_container_path = cortex_path / "engines";
 
   if (!std::filesystem::exists(engines_container_path)) {