vec-inf 0.3.3

Efficient LLM inference on Slurm clusters using vLLM.

Vector Inference: Easy inference on Slurm clusters

This repository provides an easy-to-use solution to run inference servers on Slurm-managed computing clusters using vLLM. All scripts in this repository runs natively on the Vector Institute cluster environment. To adapt to other environments, update launch_server.sh, vllm.slurm, multinode_vllm.slurm and models.csv accordingly.

Installation

If you are using the Vector cluster environment, and you don't need any customization to the inference server environment, run the following to install package:

pip install vec-inf

Otherwise, we recommend using the provided Dockerfile to set up your own environment with the package

Launch an inference server

We will use the Llama 3.1 model as example, to launch an OpenAI compatible inference server for Meta-Llama-3.1-8B-Instruct, run:

vec-inf launch Meta-Llama-3.1-8B-Instruct

You should see an output like the following:

The model would be launched using the default parameters, you can override these values by providing additional options, use --help to see the full list. You can also launch your own customized model as long as the model architecture is supported by vLLM, you'll need to specify all model launching related options to run a successful run.

You can check the inference server status by providing the Slurm job ID to the status command:

vec-inf status 13014393

You should see an output like the following:

There are 5 possible states:

• PENDING: Job submitted to Slurm, but not executed yet. Job pending reason will be shown.
• LAUNCHING: Job is running but the server is not ready yet.
• READY: Inference server running and ready to take requests.
• FAILED: Inference server in an unhealthy state. Job failed reason will be shown.
• SHUTDOWN: Inference server is shutdown/cancelled.

Note that the base URL is only available when model is in READY state, and if you've changed the Slurm log directory path, you also need to specify it when using the status command.

Finally, when you're finished using a model, you can shut it down by providing the Slurm job ID:

vec-inf shutdown 13014393

> Shutting down model with Slurm Job ID: 13014393

You call view the full list of available models by running the list command:

vec-inf list

You can also view the default setup for a specific supported model by providing the model name, for example Meta-Llama-3.1-70B-Instruct:

vec-inf list Meta-Llama-3.1-70B-Instruct

launch, list, and status command supports --json-mode, where the command output would be structured as a JSON string.

Send inference requests

Once the inference server is ready, you can start sending in inference requests. We provide example scripts for sending inference requests in examples folder. Make sure to update the model server URL and the model weights location in the scripts. For example, you can run python examples/inference/llm/completions.py, and you should expect to see an output like the following:

{"id":"cmpl-c08d8946224747af9cce9f4d9f36ceb3","object":"text_completion","created":1725394970,"model":"Meta-Llama-3.1-8B-Instruct","choices":[{"index":0,"text":" is a question that many people may wonder. The answer is, of course, Ottawa. But if","logprobs":null,"finish_reason":"length","stop_reason":null}],"usage":{"prompt_tokens":8,"total_tokens":28,"completion_tokens":20}}

NOTE: For multimodal models, currently only ChatCompletion is available, and only one image can be provided for each prompt.

SSH tunnel from your local device

If you want to run inference from your local device, you can open a SSH tunnel to your cluster environment like the following:

ssh -L 8081:172.17.8.29:8081 [e-mail] -N

Where the last number in the URL is the GPU number (gpu029 in this case). The example provided above is for the vector cluster, change the variables accordingly for your environment