Last Updated: March 27, 2025
This guide is available as a Recipe!

​
Introduction

Image captioning and labeling plays an important role in many AI and ML training workloads, and until fairly recently, has been limited in effectiveness both by available technology and cost. This guide will show you how to deploy a Vision-Language Model (VLM) for image captioning using SaladCloud. Vision-Language models provide substantial improvements over previous-generation solutions based on CLIP and BLIP. The ability to include a text prompt along with your image gives you a great deal of control as to the style and content of the returned captions. For the model, we will be using Qwen 2.5 VL 7B Instruct, an Apache 2.0 licensed model from Alibaba that excels at visual understanding, including reading text. We will use 🤗 Text Generation Inference (TGI) as an inference server. Any TGI-compatible VLM can be substituted. Prompt: What is in this image? Include details.
Snowy landscape with a blue signpost

The image shows a snowy landscape, likely taken on a mountaintop or hillside. The ground is covered with patches of snow, with some bare soil and vegetation visible where the snow has melted or been pushed away. In the distance, a hazy horizon stretches toward what appears to be valleys and mountains. There is dense evergreen forest predominantly visible on the slopes, adding depth and texture to the scene. The sky is cloudy, with slashes of sunlight breaking through on the right, indicating that the sun might be setting or emerging from behind a cloud formation. A blue signpost, common for hiking trails, is tilted on the right side, suggesting directionality. The overall atmosphere is serene and remote, typical of a high-altitude or wilderness mountainous area with no visible human structures.

​
Build A Docker Image

It is actually possible to deploy this model using just the base TGI docker image, but that method will cause the model weights to be downloaded at runtime. Since SaladCloud does not bill for the time the container image downloads, but it does bill once the container starts running, we can save costs by building a custom Docker image with the model weights pre-downloaded. First, we’re going to download the model weights and configuration files. We will do this using the TGI docker image, and mounting a local directory to /data in the container. 
docker run -it --rm --name tgi-downloader \
--env 'MODEL_ID=Qwen/Qwen2.5-VL-7B-Instruct' \
--env 'PORT=3000' \
-p 3000:3000 \
-v $(pwd)/data:/data \
--gpus all \
ghcr.io/huggingface/text-generation-inference:3.2.1
By downloading the model weights outside of the docker build, we can avoid the need to re-download the weights any time we want to change something else in our docker image. The model is quite large, and this download will take some time. Once it completes, TGI should start with the model loaded, assuming you are developing on a machine with an adequate GPU. If you are developing on a machine without a GPU, you can still download the model weights, but the server likely won’t start locally. Next, we will create a Dockerfile to build a custom image with the model weights pre-downloaded. Create a new file called Dockerfile in the same directory as the data directory, and add the following content: 
FROM ghcr.io/huggingface/text-generation-inference:3.2.1

# Copy the model weights and configuration files
COPY data /data
ENV MODEL_ID="Qwen/Qwen2.5-VL-7B-Instruct"
ENV PORT=3000
Now, build the Docker image, changing the image name and tag to suit your needs: 
docker build -t saladtechnologies/text-generation-inference:3.2.1-qwen2.5-vl-7b-instruct .
This will also take some time, as the model weights are quite large. Once it completes, you can push the image to a container registry of your choice. 
docker push saladtechnologies/text-generation-inference:3.2.1-qwen2.5-vl-7b-instruct

​
Deploy To SaladCloud

You can deploy your container group either using the Portal or the SaladCloud API. Here is an example of a container group configuration that you can use to deploy this: 
{
  "name": "tgi-qwen2-5-vl-7b-instruct",
  "display_name": "tgi-qwen2-5-vl-7b-instruct",
  "container": {
    "image": "saladtechnologies/text-generation-inference:3.2.0-qwen-2.5-vl-7b-instruct",
    "resources": {
      "cpu": 4,
      "memory": 30720,
      "gpu_classes": ["a5db5c50-cbcb-4596-ae80-6a0c8090d80f"]
    },
    "command": [],
    "priority": "high",
    "environment_variables": {
      "HOSTNAME": "::"
    },
    "image_caching": true
  },
  "autostart_policy": true,
  "restart_policy": "always",
  "replicas": 3,
  "networking": {
    "protocol": "http",
    "port": 3000,
    "auth": false,
    "load_balancer": "least_number_of_connections",
    "single_connection_limit": false,
    "client_request_timeout": 100000,
    "server_response_timeout": 100000
  },
  "startup_probe": {
    "http": {
      "path": "/health",
      "port": 3000,
      "scheme": "http",
      "headers": []
    },
    "initial_delay_seconds": 0,
    "period_seconds": 3,
    "timeout_seconds": 10,
    "success_threshold": 1,
    "failure_threshold": 50
  }
}
This configuration will deploy three replicas of the TGI server, each with 4 CPUs, 30GB of memory, and an RTX 3090 with 24GB of VRAM. The server will be accessible via HTTP on port 3000, and will be load balanced using the least number of connections algorithm. Of particular note, the environment variable HOSTNAME is set to ::, which allows the server to listen on ipv6 interfaces, as required by SaladCloud. The above example does not enable authentication, but you can by setting .networking.auth to true. Save the above configuration to a file container-group.json, and submit it to the SaladCloud API to deploy your container group. 
curl --request POST \
  --url "https://api.salad.com/api/public/organizations/${organization_name}/projects/${project_name}/containers" \
  --header 'Content-Type: application/json' \
  --header 'Salad-Api-Key: <api-key>' \
  --data @container-group.json
It will take some time for your container group to become “running.” It first must pull your container image into our internal cache, and then download the image to three compatible nodes. Once the container group is running, you can access the TGI server at the “Access Domain Name” provided in the response from the API, or via the SaladCloud Portal.

​
Using The Model

Once the container group is running, you can access swagger documentation at /docs. This will show you the available API endpoints and how to interact with them. We will be using the OpenAI-compatible /v1/chat/completions endpoint to generate image captions. To generate a caption for an image, it needs to be downloadable via a URL. This can be accomplished with just about any cloud storage provider, and can also be done with Salad’s S4 service. Here is an example of how to generate a caption for an image using the TGI server: 
curl -X 'POST' \
  'https://some-random-prefix.salad.cloud/v1/chat/completions' \
  -H 'accept: application/json' \
  -H 'Content-Type: application/json' \
  -d '{
  "max_tokens": 256,
  "messages":[
        {
            "role": "user",
            "content": [
                {
                    "type": "text",
                    "text": "What is in this image? Include details."
                },
                {
                    "type": "image_url",
                    "image_url": {
                        "url": "https://salad-benchmark-assets.download/coco2017/train2017/000000000094.jpg"
                    }
                }
            ]
        }
    ],
  "stream": false
}'
This will return a JSON response with the generated caption. You can adjust the max_tokens parameter to control the length of the generated caption. The image_url parameter should be a URL to the image you want to generate a caption for.

​
Conclusion

In this guide, we have shown you how to deploy a Vision-Language Model for image captioning using SaladCloud. This model provides state-of-the-art performance in image captioning tasks, and can be easily deployed using SaladCloud. By building a custom Docker image with the model weights pre-downloaded, you can save costs and improve performance. We have also shown you how to interact with the model using the TGI server’s API endpoints.