Distributed Workloads

Examples

• Fine-Tune LLMs with Ray and DeepSpeed on OpenShift AI
• Fine-Tune Stable Diffusion with DreamBooth and Ray Train
• Hyperparameters Optimization with Ray Tune on OpenShift AI

Integration Tests

Prerequisites

• Admin access to an OpenShift cluster (CRC is fine)

• Installed OpenDataHub or RHOAI, enabled all Distributed Workload components

• Installed Go 1.21

Common environment variables

• TEST_OUTPUT_DIR - Output directory for test logs

• TEST_TIMEOUT_SHORT - Timeout duration for short tasks

• TEST_TIMEOUT_MEDIUM - Timeout duration for medium tasks

• TEST_TIMEOUT_LONG - Timeout duration for long tasks

• TEST_RAY_IMAGE (Optional) - Ray image used for raycluster configuration

• MINIO_CLI_IMAGE (Optional) - Minio CLI image used for uploading/downloading data from/into s3 bucket

• TEST_TIER (Optional) - Specifies test tier to run, skipping tests which don’t belong to specified test tier. Supported test tiers: Smoke, Tier1, Tier2, Tier3, Pre-Upgrade and Post-Upgrade. Test tier can also be provided using test parameter testTier.

  NOTE: quay.io/modh/ray:2.47.1-py312-cu128 is the default image used for creating a RayCluster resource. If you have your own custom ray image which suits your purposes, specify it in TEST_RAY_IMAGE environment variable.

Environment variables for fms-hf-tuning test suite

• FMS_HF_TUNING_IMAGE - Image tag used in PyTorchJob CR for model training

Environment variables for fms-hf-tuning GPU test suite

• TEST_NAMESPACE_NAME (Optional) - Existing namespace where will the Training operator GPU tests be executed
• HF_TOKEN - HuggingFace token used to pull models which has limited access
• GPTQ_MODEL_PVC_NAME - Name of PersistenceVolumeClaim containing downloaded GPTQ models

To upload trained model into S3 compatible storage, use the environment variables mentioned below :

• AWS_DEFAULT_ENDPOINT - Storage bucket endpoint to upload trained dataset to, if set then test will upload model into s3 bucket
• AWS_ACCESS_KEY_ID - Storage bucket access key
• AWS_SECRET_ACCESS_KEY - Storage bucket secret key
• AWS_STORAGE_BUCKET - Storage bucket name
• AWS_STORAGE_BUCKET_MODEL_PATH (Optional) - Path in the storage bucket where trained model will be stored to

Environment variables for ODH integration test suite

• ODH_NAMESPACE - Namespace where ODH components are installed to
• NOTEBOOK_USER_NAME - Username of user used for running Workbench
• NOTEBOOK_USER_PASSWORD - Password of user used for running Workbench
• NOTEBOOK_USER_TOKEN - Login token of user used for running Workbench
• NOTEBOOK_IMAGE - Image used for running Workbench

To download MNIST training script datasets from S3 compatible storage, use the environment variables mentioned below :

• AWS_DEFAULT_ENDPOINT - Storage bucket endpoint from which to download MNIST datasets
• AWS_ACCESS_KEY_ID - Storage bucket access key
• AWS_SECRET_ACCESS_KEY - Storage bucket secret key
• AWS_STORAGE_BUCKET - Storage bucket name
• AWS_STORAGE_BUCKET_MNIST_DIR - Storage bucket directory from which to download MNIST datasets.

Environment variables for Kubeflow Trainer v2 (TrainingHub SDK) test suite

• NOTEBOOK_USER_NAME - Username of user used for running Workbench
• NOTEBOOK_USER_PASSWORD - Password of user used for running Workbench
• NOTEBOOK_IMAGE - Image used for running Workbench

To download models and datasets from S3 compatible storage for TrainingHub SDK tests, use the environment variables mentioned below :

• AWS_DEFAULT_ENDPOINT - Storage bucket endpoint
• AWS_ACCESS_KEY_ID - Storage bucket access key
• AWS_SECRET_ACCESS_KEY - Storage bucket secret key
• AWS_STORAGE_BUCKET - Storage bucket name
• AWS_STORAGE_BUCKET_SFT_DIR (Optional) - S3 prefix for SFT test data (Qwen model + Table-GPT dataset). When not set, notebooks fall back to downloading from HuggingFace.
• AWS_STORAGE_BUCKET_OSFT_DIR (Optional) - S3 prefix for OSFT test data (Qwen model + Table-GPT dataset). When not set, notebooks fall back to downloading from HuggingFace.
• AWS_STORAGE_BUCKET_LORA_DIR (Optional) - S3 prefix for LoRA test data (Qwen model + Unsloth 4-bit variant + sql-create-context dataset). When not set, notebooks fall back to downloading from HuggingFace.

See tests/trainer/resources/disconnected_env/README.md for instructions on pre-staging S3 data for disconnected environments.

Environment variables for Training Hub on Ray (RayJob) test suite

• TEST_RAY_TRAINING_HUB_IMAGE (Optional) - Ray image with Training Hub pre-installed, used for RayJob cluster configuration. Defaults to quay.io/modh/ray:2.55.1-py312-cu129-th081 (pinned by digest).

To download models and datasets from S3 compatible storage for Training Hub RayJob tests, use the environment variables mentioned below:

• AWS_DEFAULT_ENDPOINT - Storage bucket endpoint
• AWS_ACCESS_KEY_ID - Storage bucket access key
• AWS_SECRET_ACCESS_KEY - Storage bucket secret key
• AWS_STORAGE_BUCKET - Storage bucket name
• AWS_STORAGE_BUCKET_RAY_TRAINING_HUB_DIR (Optional) - S3 prefix for Training Hub RayJob test data (Qwen model + GRPO dataset subset). When not set, notebooks fall back to downloading from HuggingFace.

AI Agents

AI agents configuration lives in the ai/ directory as the single source of truth. Agent-specific config files are generated from it.

Structure

• ai/skills//SKILL.md – skill body (markdown)
• ai/skills//metadata.json – skill metadata (description, globs, etc.)
• ai/rules//RULE.md – rule body (markdown)
• ai/rules//metadata.json – rule metadata (description, globs for path-scoping)

Adding or updating a skill/rule

1. Create or edit files under ai/skills// or ai/rules//.
2. Run make sync-agents-config to regenerate agents directories e.g. .claude/.
3. Commit both the ai/ source files and the generated output together.

Note: make sync-agents-config requires uv to be installed. CI runs make verify-agents-config to ensure generated files stay in sync.

Running Tests

Execute tests like standard Go unit tests.

go test -timeout 60m ./tests/kfto/