llama-stack-provider-kft

Llama Stack Remote Post Training Provider for Distributed InstructLab Training using the Kubeflow Trainer

Utilities

As a part of this package, the ilab-kft command line interface is available to get a cluster properly set up for distributed training.

NOTE: the oc cli is a pre-requisite to using this tool.

How to

Upload your SDG data to the cluster

Currently ilab-kft allows you to upload a local directory of data to your cluster mounted in a PVC.

example: python3.11 ilab-kft.py data-upload --data-path ~/.local/share/instructlab/datasets/ --pvc-name data --namespace default

Upload your model

Using the same data-upload command, you can also upload models:

example: python3.11 ilab-kft.py data-upload --data-path ~/.cache/instructlab/models/granite-7b-lab/ --pvc-name model --namespace default

Run training

Using llama-stack and the client SDK, one can spin up a llama stack server and run post-training using this provider

llama stack run run.yaml --image-type venv

python3.10 train.py

train.py utilizes the llama-stack-client python SDK to initialize training arguments, and pass the required arguments to supervised_fine_tune in order to kick off the provider implementation maintained externally in this repository.

---

Run llama-stack-provider-kft in cluster

1. Deploy Kustomize manifests

Apply the kustomize manifests under base directory.

kubectl apply -k manifests/base/

2. (Optional) Access the service locally

If you want to run a client such as the train.py script locally, you can port-forward the service to your localhost.

kubectl port-forward svc/lls-provider-kft 8321:80