updated docs

csharp/src/Microsoft.ML.OnnxRuntime/Training/TrainingSession.shared.cs

@@ -568,8 +568,10 @@ public void OptimizerStep(RunOptions options)
         /// an inference model if it knows the inference graph outputs. The input inference graph outputs
         /// are used to prune the eval model so that the inference model's outputs align with the provided outputs.
         /// The exported model is saved at the path provided and can be used for inferencing with InferenceSession.
-        /// Note that the function re-loads the eval model from the path provided to TrainingSession
-        /// and expects that this path still be valid.
+        ///
+        /// This function modifies the eval graph in-place, so after this method is called, the TrainingSession can
+        /// no longer be used for training. In order to continue training from this point, save the checkpoint state
+        /// and create a new TrainingSession with the saved checkpoint state.
         /// </summary>
         /// <param name="inferenceModelPath">Path where the inference model should be serialized to.</param>
         /// <param name="graphOutputNames">Names of the outputs that are needed in the inference model.</param>

java/src/main/java/ai/onnxruntime/OrtTrainingSession.java

@@ -998,8 +998,9 @@ private native void schedulerStep(long apiHandle, long trainingApiHandle, long n
    * Exports the evaluation model as a model suitable for inference, setting the desired nodes as
    * output nodes.
    *
-   * <p>Note that this method reloads the evaluation model from the path provided to the training
-   * session, and this path must still be valid.
+   * <p>Note that this method modifies the eval session in-place; thus, after this method is called, the
+   * OrtTrainingSession can no longer be trained with. To continue training from this point, save the checkpoint
+   * and then load it into a new OrtTrainingSession.
    *
    * @param outputPath The path to write out the inference model.
    * @param outputNames The names of the output nodes.

objectivec/include/ort_training_session.h

@@ -229,10 +229,11 @@ NS_ASSUME_NONNULL_BEGIN
  *
  * If the training session was provided with an eval model, the training session can generate an inference model if it
  * knows the inference graph outputs. The input inference graph outputs are used to prune the eval model so that the
- * inference model's outputs align with the provided outputs. The exported model is saved at the path provided and
- * can be used for inferencing with `ORTSession`.
+ * inference model's outputs align with the provided outputs.
  *
- * @note The method reloads the eval model from the path provided to the initializer and expects this path to be valid.
+ * @note This method modifies the eval model graph in-place, so after this method is called, the ORTTrainingSession
+ * can no longer be used for training. To resume training from this point, save the checkpoint state and create a new
+ * ORTTrainingSession with the saved checkpoint state.
  *
  * @param inferenceModelPath The path to the serialized the inference model.
  * @param graphOutputNames The names of the outputs that are needed in the inference model.

orttraining/orttraining/python/training/api/module.py

@@ -194,9 +194,12 @@ def export_model_for_inferencing(
         Once training is complete, this function can be used to drop the training specific nodes in the onnx model.
         In particular, this function does the following:
 
-        - Parse over the training graph and identify nodes that generate the given output names.
+        - Parse over the eval graph and identify nodes that generate the given output names.
         - Drop all subsequent nodes in the graph since they are not relevant to the inference graph.
 
+        Once this method is called, training is considered complete and the module can no longer be used for training.
+        To resume training from this point, save the checkpoint and create a new module from the checkpoint.
+
         Args:
             inference_model_uri: The path to the inference model.
             graph_output_names: The list of output names that are required for inferencing.

orttraining/orttraining/training_api/include/onnxruntime_training_c_api.h

@@ -513,8 +513,9 @@ struct OrtTrainingApi {
    * an inference model if it knows the inference graph outputs. The input inference graph outputs
    * are used to prune the eval model so that the inference model's outputs align with the provided outputs.
    * The exported model is saved at the path provided and can be used for inferencing with InferenceSession.
-   * \note Note that the function re-loads the eval model from the path provided to OrtTrainingApi::CreateTrainingSession
-   * and expects that this path still be valid.
+   * \note Note that the function modifies the eval model graph in-place, so after this method is called, the
+   * OrtTrainingSession can no longer be used for training. To resume training from this point, save the checkpoint
+   * state and create a new OrtTrainingSession with the updated eval model.
    *
    * \param[in] sess The `this` pointer to the training session.
    * \param[in] inference_model_path Path where the inference model should be serialized to.

orttraining/orttraining/training_api/include/onnxruntime_training_cxx_api.h

@@ -336,8 +336,9 @@ class TrainingSession : public detail::Base<OrtTrainingSession> {
    * an inference model if it knows the inference graph outputs. The input inference graph outputs
    * are used to prune the eval model so that the inference model's outputs align with the provided outputs.
    * The exported model is saved at the path provided and can be used for inferencing with Ort::Session.
-   * \note Note that the function re-loads the eval model from the path provided to Ort::TrainingSession
-   * and expects that this path still be valid.
+   * \note Note that the function modifies the eval model graph in-place, so after this method is called, the
+   * OrtTrainingSession can no longer be used for training. To resume training from this point, save the checkpoint
+   * state and create a new OrtTrainingSession with the updated eval model.
    *
    * \param[in] inference_model_path Path where the inference model should be serialized to.
    * \param[in] graph_output_names Names of the outputs that are needed in the inference model.