Add support for sharing an ORT session

For every instance in a model instance group a new ORT session is
created. This code adds support to share a session per instance
group.
This support can be enabled by defining 'share_session' to true
in triton model config "parameters". Example:
parameters [
.....
  {
    key: "share_session"
    value: {string_value: "true"}
  }
]

This is a global parameter and cannot be defined per instance
group. The user should determine if the parameter makes sense for
their setup.

Author: samgoel01

src/onnxruntime.cc

@@ -25,7 +25,6 @@
 // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #include <stdint.h>
-
 #include <mutex>
 #include <vector>
 
@@ -81,10 +80,10 @@ class ModelState : public BackendModel {
   // onnx file, return in 'session' and 'allocator' the ORT session
   // and allocator.
   TRITONSERVER_Error* LoadModel(
-      const std::string& artifact_name,
+      const std::string& artifact_name, const std::string& instance_name,
       const TRITONSERVER_InstanceGroupKind instance_group_kind,
       const int32_t instance_group_device_id, std::string* model_path,
-      OrtSession** session, OrtAllocator** default_allocator,
+      std::shared_ptr<OrtSession>& session, OrtAllocator** default_allocator,
       cudaStream_t stream);
 
   const std::map<std::string, std::pair<int64_t, int64_t>>& ModelOutputs()
@@ -101,6 +100,11 @@ class ModelState : public BackendModel {
   TRITONSERVER_Error* AutoCompleteIO(
       const char* key, const OnnxTensorInfoMap& io_infos);
 
+  TRITONSERVER_Error* GetSessionForGroup(
+      const std::string& group_name, std::shared_ptr<OrtSession>& session);
+  TRITONSERVER_Error* SetSessionForGroup(
+      const std::string& group_name, const std::shared_ptr<OrtSession>& session);
+
   // Session options used when creating a ORT session.
   std::unique_ptr<OrtSessionOptions, SessionOptionsDeleter> session_options_;
 
@@ -110,6 +114,17 @@ class ModelState : public BackendModel {
   // is specified both in the output section and state section, it indicates
   // that the backend must return the output state to the client too.
   std::map<std::string, std::pair<int64_t, int64_t>> model_outputs_;
+
+  // Indicate if an onnxrt session should be shared or not. This is a model
+  // global and applies to all instances. So, storing it in the model state
+  bool share_session_;
+
+  // maintain a map of group id to onnx_rt session. This is only useful if
+  // share_session is set to true in parameters. share_session is a global model
+  // config and the user should be careful when setting this. There is no way to
+  // set this per instance group.
+  std::unordered_map<std::string, std::shared_ptr<OrtSession>>
+      groupInstanceSessionMap_;
 };
 
 TRITONSERVER_Error*
@@ -188,7 +203,7 @@ ModelState::Create(TRITONBACKEND_Model* triton_model, ModelState** state)
 }
 
 ModelState::ModelState(TRITONBACKEND_Model* triton_model)
-    : BackendModel(triton_model)
+    : BackendModel(triton_model), share_session_(false)
 {
   // Create session options that will be cloned and used for each
   // instance when creating that instance's session.
@@ -338,20 +353,31 @@ ModelState::ModelState(TRITONBACKEND_Model* triton_model)
       }
     }
   }
-
-  // FIXME. Is it possible to share a single OrtSession across
-  // multiple instances? If so then should move loading and validation
-  // of the session to here instead of creating a session for each
-  // instance in ModelStateInstance::Create().
+  
+  // This setting will apply across multiple instance groups.
+  // If this value is set all instances within an instance group will share
+  // the ort session
+  {
+    bool share_session;
+    triton::common::TritonJson::Value params;
+    if (ModelConfig().Find("parameters", &params)) {
+      THROW_IF_BACKEND_MODEL_ERROR(TryParseModelStringParameter(
+          params, "share_session", &share_session, false));
+    }
+    share_session_ = share_session;
+  }
 }
 
 TRITONSERVER_Error*
 ModelState::LoadModel(
-    const std::string& artifact_name,
+    const std::string& artifact_name, const std::string& instance_name,
     const TRITONSERVER_InstanceGroupKind instance_group_kind,
     const int32_t instance_group_device_id, std::string* model_path,
-    OrtSession** session, OrtAllocator** default_allocator, cudaStream_t stream)
+    std::shared_ptr<OrtSession>& session, OrtAllocator** default_allocator,
+    cudaStream_t stream)
 {
+  // Get the group name for the instance
+  std::string instance_group_name(GetInstanceGroupName(Name(), instance_name));
   // Find the ONNX file that describes the model itself. If the model
   // configuration doesn't have an explicit model file specified then
   // use the default name ("model.onnx").
@@ -363,6 +389,10 @@ ModelState::LoadModel(
   *model_path = JoinPath(
       {RepositoryPath(), std::to_string(Version()), cc_model_filename});
 
+  // get default cpu allocator
+  RETURN_IF_ORT_ERROR(
+      ort_api->GetAllocatorWithDefaultOptions(default_allocator));
+
   // If the model path is a directory then the actual model is
   // <dir>/model.onnx.
   {
@@ -373,6 +403,20 @@ ModelState::LoadModel(
     }
   }
 
+  // Check is we are sharing the session. If so get the session pointer and
+  // return
+  if (share_session_) {
+    if (GetSessionForGroup(instance_group_name, session) == nullptr) {
+      LOG_MESSAGE(
+          TRITONSERVER_LOG_INFO,
+          (std::string("Reusing session for group: ") + instance_group_name)
+              .c_str());
+      // Return the session
+      return nullptr;
+    }
+    // In case of error carry on with the code
+  }
+
   {
     bool exists;
     RETURN_IF_ERROR(FileExists(*model_path, &exists));
@@ -636,12 +680,22 @@ ModelState::LoadModel(
     glock.lock();
   }
 
-  RETURN_IF_ERROR(OnnxLoader::LoadSession(
-      true /* is_path */, *model_path, soptions, session));
+  {
+    // This will be allocated by OnnxRT here but will be freed when the last
+    // instance of shared_ptr is released
+    OrtSession* session_ptr;
+    RETURN_IF_ERROR(OnnxLoader::LoadSession(
+        true /* is_path */, *model_path, soptions, &session_ptr));
 
-  // get default cpu allocator
-  RETURN_IF_ORT_ERROR(
-      ort_api->GetAllocatorWithDefaultOptions(default_allocator));
+    session = std::shared_ptr<OrtSession>(session_ptr, SessionDeleter());
+
+    if (share_session_) {
+      // The session was created fine this is not a critical error
+      LOG_IF_ERROR(
+          SetSessionForGroup(instance_group_name, session),
+          "Failed to map ort session to the group for sharing");
+    }
+  }
 
   return nullptr;  // success
 }
@@ -685,7 +739,7 @@ ModelState::AutoCompleteConfig()
 
   // Must cleanup 'session'. 'allocator' is default allocator which
   // is managed by ONNX Runtime so don't need to free/release
-  std::unique_ptr<OrtSession, SessionDeleter> session;
+  std::shared_ptr<OrtSession> session;
   OrtAllocator* default_allocator;
   std::string model_path;
   {
@@ -714,12 +768,9 @@ ModelState::AutoCompleteConfig()
       }
     }
 #endif  // TRITON_ENABLE_GPU
-
-    OrtSession* sptr = nullptr;
     RETURN_IF_ERROR(LoadModel(
-        artifact_name, kind, 0, &model_path, &sptr, &default_allocator,
-        nullptr));
-    session.reset(sptr);
+        artifact_name, "", kind, 0, &model_path,
+        session, &default_allocator, nullptr));
   }
   OnnxTensorInfoMap input_tensor_infos;
   RETURN_IF_ERROR(
@@ -881,6 +932,38 @@ ModelState::AutoCompleteIO(const char* key, const OnnxTensorInfoMap& io_infos)
   return nullptr;  // success
 }
 
+TRITONSERVER_Error*
+ModelState::GetSessionForGroup(
+    const std::string& group_name, std::shared_ptr<OrtSession>& session)
+{
+  RETURN_ERROR_IF_TRUE(
+      group_name.empty(), TRITONSERVER_ERROR_INVALID_ARG,
+      std::string("Invalid group name"));
+  {
+    std::unordered_map<std::string, std::shared_ptr<OrtSession>>::iterator
+        sessionEntry;
+    sessionEntry = groupInstanceSessionMap_.find(group_name);
+    RETURN_ERROR_IF_TRUE(
+        (sessionEntry == groupInstanceSessionMap_.end()),
+        TRITONSERVER_ERROR_NOT_FOUND, std::string("No such group"));
+
+    session = sessionEntry->second;
+  }
+  return nullptr;
+}
+
+TRITONSERVER_Error*
+ModelState::SetSessionForGroup(
+    const std::string& group_name, const std::shared_ptr<OrtSession>& session)
+{
+  RETURN_ERROR_IF_TRUE(
+      group_name.empty(), TRITONSERVER_ERROR_INVALID_ARG,
+      std::string("Invalid group name"));
+
+  groupInstanceSessionMap_[group_name] = session;
+  return nullptr;
+}
+
 //
 // ModelInstanceState
 //
@@ -967,7 +1050,7 @@ class ModelInstanceState : public BackendModelInstance {
 
   // Onnx Runtime variables that are used across runs on this
   // instance.
-  OrtSession* session_;
+  std::shared_ptr<OrtSession> session_;
   OrtAllocator* default_allocator_;
   OrtMemoryInfo* cuda_allocator_info_;
   const OrtMemoryInfo* cpu_allocator_info_;
@@ -1013,7 +1096,7 @@ ModelInstanceState::ModelInstanceState(
       io_binding_(nullptr), output_buffer_(nullptr)
 {
   THROW_IF_BACKEND_INSTANCE_ERROR(model_state->LoadModel(
-      ArtifactFilename(), Kind(), DeviceId(), &model_path_, &session_,
+      ArtifactFilename(), Name(), Kind(), DeviceId(), &model_path_, session_,
       &default_allocator_, CudaStream()));
 
   if (Kind() == TRITONSERVER_INSTANCEGROUPKIND_GPU) {
@@ -1026,7 +1109,7 @@ ModelInstanceState::ModelInstanceState(
       ort_api->AllocatorGetInfo(default_allocator_, &cpu_allocator_info_));
 
   THROW_IF_BACKEND_INSTANCE_ORT_ERROR(
-      ort_api->CreateIoBinding(session_, &io_binding_));
+      ort_api->CreateIoBinding(session_.get(), &io_binding_));
 
   THROW_IF_BACKEND_INSTANCE_ORT_ERROR(ort_api->CreateRunOptions(&runOptions_));
 
@@ -1114,9 +1197,6 @@ ModelInstanceState::~ModelInstanceState()
   ort_api->ReleaseRunOptions(runOptions_);
   ort_api->ReleaseIoBinding(io_binding_);
   ort_api->ReleaseMemoryInfo(cuda_allocator_info_);
-  if (session_ != nullptr) {
-    OnnxLoader::UnloadSession(session_);
-  }
   // 'default_allocator_' is default allocator which is managed by ONNX
   // Runtime
 }
@@ -1176,7 +1256,7 @@ ModelInstanceState::ValidateBooleanSequenceControl(
   if (*have_control) {
     OnnxTensorInfoMap input_tensor_infos;
     RETURN_IF_ERROR(
-        InputInfos(session_, default_allocator_, input_tensor_infos));
+        InputInfos(session_.get(), default_allocator_, input_tensor_infos));
     const auto& iit = input_tensor_infos.find(tensor_name);
     if (iit == input_tensor_infos.end()) {
       return TRITONSERVER_ErrorNew(
@@ -1233,7 +1313,7 @@ ModelInstanceState::ValidateTypedSequenceControl(
   if (*have_control) {
     OnnxTensorInfoMap input_tensor_infos;
     RETURN_IF_ERROR(
-        InputInfos(session_, default_allocator_, input_tensor_infos));
+        InputInfos(session_.get(), default_allocator_, input_tensor_infos));
     const auto& iit = input_tensor_infos.find(tensor_name);
     if (iit == input_tensor_infos.end()) {
       return TRITONSERVER_ErrorNew(
@@ -1280,10 +1360,11 @@ TRITONSERVER_Error*
 ModelInstanceState::ValidateInputs(const size_t expected_input_cnt)
 {
   std::set<std::string> input_tensor_names;
-  RETURN_IF_ERROR(InputNames(session_, input_tensor_names));
+  RETURN_IF_ERROR(InputNames(session_.get(), input_tensor_names));
 
   OnnxTensorInfoMap input_tensor_infos;
-  RETURN_IF_ERROR(InputInfos(session_, default_allocator_, input_tensor_infos));
+  RETURN_IF_ERROR(
+      InputInfos(session_.get(), default_allocator_, input_tensor_infos));
 
   if (input_tensor_infos.size() != expected_input_cnt) {
     return TRITONSERVER_ErrorNew(
@@ -1368,10 +1449,10 @@ TRITONSERVER_Error*
 ModelInstanceState::ValidateOutputs()
 {
   std::set<std::string> output_tensor_names;
-  RETURN_IF_ERROR(OutputNames(session_, output_tensor_names));
+  RETURN_IF_ERROR(OutputNames(session_.get(), output_tensor_names));
 
   RETURN_IF_ERROR(
-      OutputInfos(session_, default_allocator_, output_tensor_infos_));
+      OutputInfos(session_.get(), default_allocator_, output_tensor_infos_));
 
   triton::common::TritonJson::Value ios;
   RETURN_IF_ERROR(model_state_->ModelConfig().MemberAsArray("output", &ios));
@@ -1765,7 +1846,7 @@ ModelInstanceState::OrtRun(
     const uint32_t response_count)
 {
   RETURN_IF_ORT_ERROR(
-      ort_api->RunWithBinding(session_, runOptions_, io_binding_));
+      ort_api->RunWithBinding(session_.get(), runOptions_, io_binding_));
   return nullptr;
 }
 
@@ -2267,7 +2348,6 @@ ModelInstanceState::ReadOutputTensors(
         }
       }
 
-
     } else {
       char* output_buffer = nullptr;
       RETURN_IF_ORT_ERROR(

src/onnxruntime_utils.cc

@@ -493,5 +493,22 @@ CompareDimsSupported(
   return nullptr;  // success
 }
 
+std::string
+GetInstanceGroupName(
+    const std::string& model_name, const std::string& instance_name)
+{
+  std::regex groupNameRegex('(' + model_name + '_' + "[0-9]" + ')');
+  std::smatch groupName;
+
+  if (model_name.empty() || instance_name.empty()) {
+    return "";
+  }
+
+  if (std::regex_search(instance_name, groupName, groupNameRegex)) {
+    return groupName.str(1);
+  }
+
+  return "";
+}
 
 }}}  // namespace triton::backend::onnxruntime

src/onnxruntime_utils.h

@@ -27,6 +27,7 @@
 #pragma once
 
 #include <onnxruntime_c_api.h>
+#include <regex>
 #include <set>
 #include <string>
 #include <unordered_map>
@@ -149,4 +150,7 @@ TRITONSERVER_Error* CompareDimsSupported(
     const std::vector<int64_t>& model_shape, const std::vector<int64_t>& dims,
     const int max_batch_size, const bool compare_exact);
 
+std::string GetInstanceGroupName(
+    const std::string& model_name, const std::string& instance_name);
+
 }}}  // namespace triton::backend::onnxruntime