Red Hat’s field demo turns OSFT into an OpenShift AI training-to-serving path

Red Hat AI Americas has published a complete reference build for fine-tuning and serving a banking-support router on Red Hat OpenShift AI 3.4. The repository’s single initial commit landed on August 17, but the pattern is useful now because it joins the training and inference halves that are often documented separately. (repository; initial commit)

What the build does

The demo starts with a small language model that produces unreliable conversational output when asked to classify banking complaints. It fine-tunes that model with Orthogonal Subspace Fine-Tuning against the Banking77 dataset, whose 10,000 examples span 77 banking intents. The target is deliberately operational rather than conversational: strict JSON routing decisions that another application can consume. (README)

Training runs through Kubeflow Trainer v2 using the training-hub runtime. The documented topology uses two nodes with two GPUs each, writes checkpoints to a shared ReadWriteMany persistent volume and sends metrics to MLflow. A GPU-backed OpenShift AI workbench drives the notebook and requires service-account permissions to create Trainer resources and reach MLflow. (README)

The useful part is the handoff

After training, the workflow converts the checkpoint to Hugging Face format on the shared volume, removes a tokenizer quantization block that the repository says can prevent vLLM from loading the result, and deploys the model through OpenShift AI’s single-model serving interface. KServe and the vLLM ServingRuntime then expose the fine-tuned router through an OpenAI-compatible /v1 endpoint. There is no separate model-export or object-storage transfer step in the documented path. (README)

That makes the repository more than a notebook. It includes namespace, service-account, RBAC and persistent-volume manifests; a custom workbench container; setup and teardown scripts; and troubleshooting for Trainer permissions, MLflow injection, shared storage and serving failures. The prerequisites are substantial—OpenShift 4.19 or later, OpenShift AI 3.4, four training GPUs across two nodes, another GPU for the workbench and RWX storage—but they are stated rather than hidden. (README)

What platform teams should take from it

The strongest reusable idea is the contract between stages: one shared model path, explicit identities for the notebook and workers, metrics recorded during training, and a serving runtime that reads the resulting weights directly. Teams evaluating OpenShift AI’s newer training stack can use the repository as a checklist for the platform plumbing around a distributed job, even if they replace the banking dataset or OSFT method. The repository has no tagged release and only its initial commit, so it should be treated as a field reference build rather than a supported product artifact. (repository; initial commit)