Skip to main content

Get Started

This section guides you through setting up and using RAIC's Model Registry to upload and manage models for deployment.

Step 1: Create a Registry​

  1. From the main navigation, go to Model Registry.
  2. Click on the Registry tab in the sub-menu on the top.
  3. Select Create New Registry.
note

Every new organization starts with a default registry already configured. If you need a custom setup, you can create a new registry and choose your preferred location—this is where your model weights will be stored and made available for deployment to Endpoints.

Step 2: Create a New Model​

Once your registry is set up:

  1. Go to the Model sub-menu.
  2. Click Create New Model.
  3. In the form, provide:
    • A model name
    • The registry where it should reside
    • The preferred GPU types (you can select more than one)
info

Selecting the GPU types during model creation ensures that, at deployment time, the platform can match your model with compatible GPUs available in the chosen registry location.

Step 3: Install the CLI and set your API token​

Model weights are uploaded with the Radiant CLI. If you haven't already, follow the CLI installation guide to install raic and configure your RAIC_ACCESS_TOKEN. Once raic vm list-sku returns a list of SKUs, you're ready to continue.

Step 4: Upload Model Weights via CLI​

After creating a model, you’ll be taken to its details view, where you can view and manage its versions.

Download or prepare your model as a HuggingFace-format directory, then upload the whole directory with raic model upload. Use the SKU shown on the model's details page for --sku, and any version tag you like for --tag.

Worked example — uploading Qwen2.5-0.5B-Instruct to a model with SKU test-custom-model-1 in the default-registry:

huggingface-cli download Qwen/Qwen2.5-0.5B-Instruct --local-dir ./qwen2.5-0.5b-instruct
raic model upload ./qwen2.5-0.5b-instruct \
  --sku test-custom-model-1 \
  --registry-name default-registry \
  --tag v0.0.1

The CLI walks the directory and uploads each file, printing a progress bar per file and Successfully uploaded! when the version is complete.

caution

Pass the directory, not a single weights file. vLLM needs config.json and the tokenizer files alongside the .safetensors weights. A version containing only the weights file will upload successfully but fail to serve when you deploy it.

Upload Status and Version Availability​
  • Upload progress is visible on both the UI and CLI.
  • If the upload is interrupted or fails, the status will show as UploadFailed.
  • Once uploaded successfully:
    • The version status changes to Uploaded
    • Then transitions to Synchronizing
    • Finally becomes Available, and ready for deployment to Endpoints.

✨ Alternative: Use Fine-Tuned Models​

You don’t always need to create and upload a model manually.

Launching a fine-tuning job creates a placeholder model in the registry for you, and you register the job's checkpoints as versions of it — so Steps 2 to 4 above are taken care of. See Fine-Tuning Guide for the entire flow.

Contents