Helix Documentation
Private Deployment
Manual Install
Runpod

Deploying a Runpod Runner

Create a runpod GPU pod template like this:

Don’t use the version or arguments in the screenshot above, use the details below

Container image:

registry.helix.ml/helix/runner:<LATEST_TAG>

Where <LATEST_TAG> is the tag of the latest release in the form X.Y.Z from https://github.com/helixml/helix/releases

You can also use X.Y.Z-small to use an image with Llama3-8B and Phi3-Mini pre-baked (llama3:instruct,phi3:instruct), or X.Y.Z-large for one with all our supported Ollama models pre-baked. Warning: the large image is large (over 100GB), but it saves you re-downloading the weights every time the container restarts! We recommend using X.Y.Z-small and setting the RUNTIME_OLLAMA_WARMUP_MODELS environment variable to llama3:instruct,phi3:instruct to get started (in the runpod UI), so the download isn’t too big. If you want to use other models in the interface, don’t specify RUNTIME_OLLAMA_WARMUP_MODELS environment variable, and it will use the defaults (all models).

Docker Command:

--api-host https://<YOUR_CONTROLPLANE_HOSTNAME> --api-token <RUNNER_TOKEN_FROM_ENV> --runner-id runpod-001 --memory <GPU_MEMORY>GB --allow-multiple-copies

Replace <RUNNER_TOKEN_FROM_ENV> and <GPU_MEMORY> accordingly. You might want to update the runner-id with a more descriptive name, and make sure it’s unique. That ID will show up in the helix dashboard at https://<YOUR_CONTROLPLANE_HOSTNAME>/dashboard for admin users.

Then start runners from your template, customizing the docker command accordingly.

Configuring a Runner

  • You can update RUNTIME_OLLAMA_WARMUP_MODELS to match the specific Ollama models you want to enable for your Helix install, see available values.
  • Helix will download the weights for models specified in RUNTIME_OLLAMA_WARMUP_MODELS at startup if they are not baked into the image. This can be slow, especially if it runs in parallel across many runners, and can easily saturate your network connection. This is why using the images with pre-baked weights (-small and -large variants) is recommended.
  • Warning: the -large image is large (over 100GB), but it saves you re-downloading the weights every time the container restarts! We recommend using X.Y.Z-small and setting the RUNTIME_OLLAMA_WARMUP_MODELS value to llama3:instruct,phi3:instruct to get started, so the download isn’t too big. If you want to use other models in the Helix UI and API, delete this -e RUNTIME_OLLAMA_WARMUP_MODELS line from below, and it will use the defaults (all models). The default models will take a long time to download!
  • Update <GPU_MEMORY> to correspond to how much GPU memory you have, e.g. “80GB” or “24GB”
  • You can add --gpus 1 before the image name to target a specific GPU on the system (starting at 0). If you want to use multiple GPUs on a node, you’ll need to run multiple runner containers (in that case, remember to give them different names)
  • Make sure to run the container with --restart always or equivalent in your container runtime, since the runner will exit if it detects an unrecoverable error and should be restarted automatically
  • If you want to run the runner on the same machine as the controlplane, either: (a) set --network host and set --api-host http://localhost:8080 so that the runner can connect on localhost via the exposed port, or (b) use --api-host http://172.17.0.1:8080 so that the runner can connect to the API server via the docker bridge IP. On Windows or Mac, you can use --api-host http://host.docker.internal:8080
  • Helix will currently also download and run SDXL and Mistral-7B weights used for fine-tuning at startup. These weights are not currently pre-baked anywhere. This can be disabled with RUNTIME_AXOLOTL_ENABLED=false if desired. If running in a low-memory environment, this may cause CUDA OOM errors at startup, which can be ignored (at startup) since the scheduler will only fit models into available memory after the startup phase.
  • If you want to use text fine-tuning, you need to set the environment variable HF_TOKEN to a valid Huggingface token, then you now need to accept sharing your contact information with Mistral here and then fetch an access token from here and then specify it in this environment variable.
Last updated on
Kubernetes - Helm