Loading...
Loading...
Workflow classification predictions may fail with an “Unable to predict” message. This affects the Machine Learning Engine (MLE) ML Predictor and ML Trainer components for the workflow_classification_trainer capability. All versions are affected. The issue impacts models trained during 8–14 October 2026, but the training date alone does not confirm that a model is affected. Confirm the error signature before taking action. The failure has two possible causes. First, a training-service dependency update exported models in ONNX IR version 10, which the Predictor cannot load. The Predictor returns “Unsupported model IR version: 10, max supported IR version: 8.” Second, for some neural-network models, the weights were saved in a separate .onnx.data file that was not persisted during upload, so the model cannot load. Models with missing weights cannot be repaired and must be retrained. The corrective release MLE v2026.10.1 has already been deployed to Production backend service and automatically applies to your instance.
Run a prediction over the workflow_classification_trainer capability on the solutions trained between 8th and 14th Oct The following error will show: "Unable to predict"
MLE v2026.10.1 includes an updated prediction runtime and model-export corrections. The runtime update addresses the ONNX IR compatibility issue, and the export corrections package weights into a single model file for the affected classifiers. Before retraining, confirm the error: If the error is an unsupported IR version: Validate predictions after the corrected runtime is available. If the error persists, contact ServiceNow Support to confirm the runtime version and model compatibility. If the error references a missing classifier.onnx.data file: Retrain the affected classification solution after the corrected training service is available. The runtime update alone does not restore missing model weights. Select one of the retraining methods below. Submitting the same training job through both methods is not required. Method 1 - retrain using the user interface: Open the All navigation menu. Navigate to Classification > Classification Definitions. Locate and open the affected classification definition. Select Update and Retrain. Monitor the training job using the steps outlined in the below section "Monitor Training and Verify Recovery". Method 2 - Submit training using Scripts-Background: Testing should be performed in a non-production instance first. This method requires the necessary administrative access and authorization to run background scripts (All > System Definition > Scripts - Background). Run the attached script (PRB2099596_retrain_script.txt) in the solution’s application scope. Replace REPLACE_WITH_SOLUTION_NAME with the solution’s Name, not its display label or sys_id. Submitting the script starts a training job; it does not confirm that training has completed or that predictions have recovered. Monitor Training and Verify Recovery: In the navigation filter, enter ml_solution.list to open the ML Solutions table. Locate the affected solution and review the training progress. Confirm that the new training job completes successfully. Run a prediction using representative input for the affected solution. Confirm that a prediction is returned and that “Unable to predict” no longer appears. If prediction-service logs are available, confirm that the relevant ONNX model-loading error no longer occurs.
PRB2099596
Click on a version to see all relevant bugs
ServiceNow Integration
Learn more about where this data comes from
BugZero Plan
Streamline upgrades with automated vendor bug scrubs
BugZero Prevent
Wish you caught this bug sooner? Get proactive today.