Implementation:Microsoft Onnxruntime OrtTrainingApis

Knowledge Sources	Microsoft_Onnxruntime
Domains	Training, API, C_API
Last Updated	2026-02-10 04:00 GMT

Overview

Declares the complete set of ORT Training C API function signatures that provide the public-facing interface for training session management, checkpoint operations, and parameter manipulation.

Description

The `ort_training_apis.h` header defines the `OrtTrainingApis` namespace containing all C API entry points for the ORT Training functionality. These APIs are consumed by language bindings (Python, C#, Java) and provide a stable ABI. The functions cover the full training lifecycle:

• Session Management: `CreateTrainingSession` and `CreateTrainingSessionFromBuffer` create training sessions from file paths or in-memory byte arrays. `ReleaseTrainingSession` handles cleanup.
• Training Execution: `TrainStep` and `EvalStep` execute forward/backward and forward-only passes respectively. `LazyResetGrad` resets gradients lazily. `OptimizerStep` performs parameter updates.
• Learning Rate: `SetLearningRate`, `GetLearningRate`, `RegisterLinearLRScheduler`, and `SchedulerStep` manage learning rate schedules.
• Checkpoint: `LoadCheckpoint`, `LoadCheckpointFromBuffer`, `SaveCheckpoint`, and `ReleaseCheckpointState` handle checkpoint persistence.
• Parameters: `GetParametersSize`, `CopyParametersToBuffer`, `CopyBufferToParameters`, `GetParameterTypeAndShape`, `UpdateParameter`, and `GetParameter` enable parameter introspection and manipulation.
• Properties: `AddProperty` and `GetProperty` manage user-defined checkpoint properties.
• Model Export: `ExportModelForInferencing` exports a trained model for deployment.
• Utilities: `SetSeed` sets the random seed, `GetTrainingApi` returns the API struct for a given version.

Usage

Use this header when implementing C API bindings or when calling the ORT Training API directly from C/C++ code. This is the primary entry point for all training operations exposed through the ORT C API.

Code Reference

Source Location

• Repository: Microsoft_Onnxruntime
• File: orttraining/orttraining/training_api/ort_training_apis.h
• Lines: 1-107

Signature

namespace OrtTrainingApis {

ORT_API(const OrtTrainingApi*, GetTrainingApi, uint32_t version);

ORT_API_STATUS_IMPL(CreateTrainingSession, _In_ const OrtEnv* env,
                    _In_ const OrtSessionOptions* options,
                    _Inout_ OrtCheckpointState* checkpoint_state,
                    _In_ const ORTCHAR_T* train_model_path,
                    _In_ const ORTCHAR_T* eval_model_path,
                    _In_ const ORTCHAR_T* optimizer_model_path,
                    _Outptr_result_maybenull_ OrtTrainingSession** out);

ORT_API_STATUS_IMPL(TrainStep, _Inout_ OrtTrainingSession* session,
                    _In_opt_ const OrtRunOptions* run_options,
                    _In_ size_t inputs_len,
                    _In_reads_(inputs_len) const OrtValue* const* inputs,
                    _In_ size_t outputs_len,
                    _Inout_updates_all_(outputs_len) OrtValue** outputs);

ORT_API_STATUS_IMPL(EvalStep, _In_ const OrtTrainingSession* session,
                    _In_opt_ const OrtRunOptions* run_options,
                    _In_ size_t inputs_len,
                    _In_reads_(inputs_len) const OrtValue* const* inputs,
                    _In_ size_t outputs_len,
                    _Inout_updates_all_(outputs_len) OrtValue** outputs);

ORT_API_STATUS_IMPL(OptimizerStep, _Inout_ OrtTrainingSession* session,
                    _In_opt_ const OrtRunOptions* run_options);

ORT_API_STATUS_IMPL(LoadCheckpoint, _In_ const ORTCHAR_T* checkpoint_path,
                    _Outptr_ OrtCheckpointState** checkpoint_state);

ORT_API_STATUS_IMPL(SaveCheckpoint, _In_ OrtCheckpointState* checkpoint_state,
                    _In_ const ORTCHAR_T* checkpoint_path,
                    const bool include_optimizer_state);

ORT_API(void, ReleaseCheckpointState, _Frees_ptr_opt_ OrtCheckpointState* checkpoint_state);
ORT_API(void, ReleaseTrainingSession, _Frees_ptr_opt_ OrtTrainingSession* session);

}  // namespace OrtTrainingApis

Import

#include "orttraining/training_api/ort_training_apis.h"

I/O Contract

API Function	Key Inputs	Key Outputs	Description
CreateTrainingSession	OrtEnv, OrtSessionOptions, checkpoint, model paths	OrtTrainingSession*	Creates a new training session from model files
TrainStep	OrtTrainingSession, inputs array	outputs array, OrtStatus	Executes forward+backward pass
EvalStep	OrtTrainingSession, inputs array	outputs array, OrtStatus	Executes forward-only evaluation pass
OptimizerStep	OrtTrainingSession	OrtStatus	Performs one optimizer update step
LoadCheckpoint	checkpoint path	OrtCheckpointState*	Loads checkpoint from file
SaveCheckpoint	OrtCheckpointState, path, bool	OrtStatus	Saves checkpoint to file
GetParameter	OrtCheckpointState, name	OrtValue*	Retrieves a named parameter value
UpdateParameter	OrtCheckpointState, name, OrtValue	OrtStatus	Updates a named parameter value

Usage Examples

// C API usage pattern
OrtCheckpointState* ckpt = nullptr;
OrtTrainingApis::LoadCheckpoint(ORT_TSTR("checkpoint.ckpt"), &ckpt);

OrtTrainingSession* session = nullptr;
OrtTrainingApis::CreateTrainingSession(env, options, ckpt,
    ORT_TSTR("train.onnx"), ORT_TSTR("eval.onnx"),
    ORT_TSTR("optimizer.onnx"), &session);

// Run training step
OrtTrainingApis::TrainStep(session, nullptr, num_inputs, inputs, num_outputs, outputs);
OrtTrainingApis::OptimizerStep(session, nullptr);

// Cleanup
OrtTrainingApis::ReleaseTrainingSession(session);
OrtTrainingApis::ReleaseCheckpointState(ckpt);

Related Pages

• Implementation:Microsoft_Onnxruntime_TrainingSession
• Implementation:Microsoft_Onnxruntime_Checkpoint_API
• Implementation:Microsoft_Onnxruntime_TrainingModule
• Implementation:Microsoft_Onnxruntime_TrainingOptimizer