示例

• 在OpenShift AI上用Ray和DeepSpeed微调大型语言模型
• 与DreamBooth和Ray Train合作的微调稳定扩散
• OpenShift AI 上的 Ray Tune 超参数优化

积分测试

先决条件

• 管理员访问OpenShift集群（CRC可以）

• 安装了OpenDataHub或RHOAI，启用了所有分布式工作负载组件

• 安装了Go 1.21

常见环境变量

• TEST_OUTPUT_DIR - 测试日志输出目录
• TEST_TIMEOUT_SHORT - 短任务的超时时间
• TEST_TIMEOUT_MEDIUM - 中等任务的超时时间
• TEST_TIMEOUT_LONG - 长时间任务的超时时间
• TEST_RAY_IMAGE（可选）- 用于光线集群配置的光线图像
• MINIO_CLI_IMAGE（可选）- 用于上传/下载数据到S3桶的Minio CLI镜像
• TEST_TIER（可选）- 指定运行的测试等级，跳过不属于指定测试等级的测试。支持的测试等级：烟雾、第一层、第二层、第三层、升级前和升级后。测试层也可以通过测试参数 testTier提供。

注意：quay.io/modh/ray:2.47.1-py312-cu128 是用于创建 RayCluster 资源的默认镜像。

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