Add a model master from a vLLM recipe
Create a model master from a model's published vLLM recipe, instead of entering the serving values manually.
Create a model master from a model's published vLLM recipe, instead of entering the serving values manually. You can do this for any model that vLLM supports, even when CosmicAC ships no parameters for that model.
Prerequisites
You need the following before you start:
- A running CosmicAC deployment. See Installation.
- Access to your CosmicAC web interface.
- A published vLLM recipe for the model. To check, open
https://recipes.vllm.ai/<namespace>/<name>, such as Qwen/Qwen3.8-Flash-Next.
Steps
Open the Register model page
In the navigation menu, click Models, and then click Register model.
Select the model type
Select Deployable · vLLM.
Enter the model repository
In Model repository, enter the model's Hugging Face repository ID in the format namespace/name, such as Qwen/Qwen3-VL.
Adjust the values for your hardware
Under Recommended values, set the following fields:
- Recommended GPU count: select 1, 2, 4, 8, or 16 GPUs for one replica.
- Runtime image: select the vLLM version that serves the model.
- Root disk: enter a disk size in GB that fits the model's weights.
Set which job values trigger a warning
Under Deviation directions, select a setting for GPU count, GPU memory utilisation, Max concurrent sequences, and Max model length:
| Setting | CosmicAC warns when the new job's value is |
|---|---|
| Below unsafe | Lower than the value you set. |
| Above unsafe | Higher than the value you set. |
| Exact | Different from the value you set. |
Register the model
Click Register.
CosmicAC saves the model master and returns you to the Models page. New jobs for that model now start from these values.