Training and Serving Custom Models on SAP AI Core
As of 2026-09-25
SAP AI Core's train-and-serve capability is language-agnostic and independent of the generative AI hub: any ML framework, containerized, run as an Argo Workflows execution, with output artifacts served through a KServe-backed deployment. This module works from SAP's own service guide to walk the training workflow step by step (instance, template, scenario, configuration, execution, logs), the mandatory default object store secret without which no training pipeline can write output, the three patterns for getting a model onto the tenant (train from scratch, fine-tune or continue, package a pretrained artifact), a serving template's KServe InferenceService spec and its scaling parameters, and the efficiency features that keep an always-on endpoint affordable. It closes with artifact signatures and training schedules for auditable, automated retraining, and a triage guide mapping the three most common failures to their root object. Three exercises and a self-assessment gate the move to M364.
What you will learn
- Explain why training on SAP AI Core is an Argo-based workflow execution, not a managed 'train' button, and what a language-agnostic training contract requires from a data-science team
- Walk a training workflow end to end — choose an instance, define the template, list scenarios, create configurations, start training, retrieve logs — and name the object each step produces
- Choose between training from scratch, fine-tuning or continuing an imported model, and packaging an already-trained model for serving only, for a stated client need
- Configure a serving template's KServe InferenceService spec — predictor, STORAGE_URI, minReplicas, maxReplicas, containerConcurrency — for a real traffic pattern
- Use artifact signatures and training schedules to make repeated training runs auditable and automated retraining safe
- Diagnose the three most common training-and-serving failures from their symptoms alone: missing globalName, missing default object store secret, broken registry or artifact reference
Who this is for. M362 gave you SAP AI Core's object model and its resource-group boundaries. This module puts them to work on the case that model was built for: a client whose need is not "call an LLM through orchestration" but "train our own model on our own data, and serve it as a production endpoint" — a fraud score, a delivery-risk prediction, a demand forecast, or a fine-tuned model that a hyperscaler catalogue does not cover. This is SAP AI Core's own train-and-serve capability, language-agnostic and independent of the generative AI hub, and it is where a data-science background stops being optional.
1. Training is a workflow execution, not a button
There is no "train" button in SAP AI Core. Training is an execution — an instance of a non-deployable executable, run through the Argo Workflows engine covered in M362. SAP is explicit that AI Core is "language-agnostic for model training code": you write your training script in whatever language and framework you choose, containerize it, and let AI Core orchestrate the run; the platform does not impose scikit-learn, PyTorch or any other library. What AI Core does impose is the workflow contract around your code:
Prerequisites
- M362 (SAP AI Core — Resource Groups, Scenarios, Executables) — this module builds directly on its object model
- Working knowledge of at least one ML framework (scikit-learn, PyTorch, or similar) and comfort writing a Dockerfile
- Optional for hands-on work: a trial SAP AI Core tenant with a registered Git repository and Docker registry secret
Outcomes
- Stand up a complete training-to-serving pipeline for a custom model on SAP AI Core, from workflow template to a live inference endpoint
- Choose the right onboarding pattern (train from scratch, fine-tune, or serve-only) for a stated client model and justify it
- Size a deployment's scaling parameters against a stated traffic profile and defend the cost/latency trade-off
- Triage a broken training or serving setup to its root object (artifact, secret, registry, quota) from the symptom alone
Full module available to members. The full module adds: the decision framework · the end-to-end scenario walkthrough · the KPI scorecard · the anti-patterns · the code blocks · the knowledge check · the diagrams.