Revert "Load PyTorch backend as a persistent backend (#224)" (#250)

This reverts commit 0811529.

Author: Tabrizian

src/backend_manager.cc

@@ -381,43 +381,4 @@ TritonBackendManager::BackendState(
   return Status::Success;
 }
 
-Status
-TritonBackendManager::PreloadBackend(
-    const std::string& backend_name,
-    const triton::common::BackendCmdlineConfigMap& config_map)
-{
-  std::string backends_dir;
-  std::string specialized_backend_name;
-  std::string backend_libname;
-  RETURN_IF_ERROR(
-      BackendConfigurationGlobalBackendsDirectory(config_map, &backends_dir));
-  RETURN_IF_ERROR(BackendConfigurationSpecializeBackendName(
-      config_map, backend_name, &specialized_backend_name));
-  RETURN_IF_ERROR(BackendConfigurationBackendLibraryName(
-      specialized_backend_name, &backend_libname));
-
-  const auto backend_dir = JoinPath({backends_dir, specialized_backend_name});
-  const auto backend_libpath = JoinPath({backend_dir, backend_libname});
-  bool exists = false;
-  RETURN_IF_ERROR(FileExists(backend_libpath, &exists));
-  if (exists) {
-    triton::common::BackendCmdlineConfig empty_backend_cmdline_config;
-    const triton::common::BackendCmdlineConfig* config;
-    const auto& itr = config_map.find(backend_name);
-    if (itr == config_map.end()) {
-      config = &empty_backend_cmdline_config;
-    } else {
-      config = &itr->second;
-    }
-
-    // Backend manager would always hold a reference to the backend object
-    // so it is ok if this object goes out of scope.
-    std::shared_ptr<TritonBackend> backend;
-    RETURN_IF_ERROR(CreateBackend(
-        backend_name, backend_dir, backend_libpath, *config, &backend));
-  }
-
-  return Status::Success;
-}
-
 }}  // namespace triton::core

src/backend_manager.h

@@ -31,7 +31,6 @@
 #include <string>
 #include <unordered_map>
 
-#include "backend_config.h"
 #include "constants.h"
 #include "filesystem/api.h"
 #include "server_message.h"
@@ -170,10 +169,6 @@ class TritonBackendManager {
       const triton::common::BackendCmdlineConfig& backend_cmdline_config,
       std::shared_ptr<TritonBackend>* backend);
 
-  Status PreloadBackend(
-      const std::string& backend_name,
-      const triton::common::BackendCmdlineConfigMap& config_map);
-
   Status BackendState(
       std::unique_ptr<
           std::unordered_map<std::string, std::vector<std::string>>>*
@@ -182,7 +177,6 @@ class TritonBackendManager {
  private:
   DISALLOW_COPY_AND_ASSIGN(TritonBackendManager);
   TritonBackendManager() = default;
-
   std::unordered_map<std::string, std::shared_ptr<TritonBackend>> backend_map_;
 };
 

src/server.cc

@@ -151,17 +151,6 @@ InferenceServer::Init()
     return status;
   }
 
-  // TODO: Remove once the PyTorch bug is resolved. Currently, PyTorch has some
-  // issues with simultaneous model loading of other backends causing a segfault
-  // (TF to be specific). Once those issues are resolved we can remove this
-  // change.
-  status =
-      backend_manager_->PreloadBackend("pytorch", backend_cmdline_config_map_);
-  if (!status.IsOk()) {
-    ready_state_ = ServerReadyState::SERVER_FAILED_TO_INITIALIZE;
-    return status;
-  }
-
   // CacheManager
   status = TritonCacheManager::Create(&cache_manager_, cache_dir_);
   if (!status.IsOk()) {
@@ -249,7 +238,6 @@ InferenceServer::Init()
     LOG_WARNING << status.Message();
   }
 
-
   // Create the model manager for the repository. Unless model control
   // is disabled, all models are eagerly loaded when the manager is created.
   bool polling_enabled = (model_control_mode_ == ModelControlMode::MODE_POLL);