FAQ - AI Tools

The ovhai command line interface (CLI) is a tool provided by OVHcloud that allows users to interact with their AI Solutions (notebooks, jobs, apps, registries, data, etc.) directly from the command line. This is the recommended means of interaction with AI Notebooks, AI Training, and AI Deploy.

To install the ovhai CLI, you can follow the installation instructions provided by OVHcloud in the official ovhai installation documentation.

After completing the ovhai CLI installation, remember to connect to your AI Platform user. This will enable you to manage the notebooks, jobs, and apps associated with this user. To log in, run the command ovhai login and enter the credentials of your AI Platform user.

You will have to use ovhai login with your AI Platform user credentials to authenticate yourself. The --token parameter has been created for AI Solutions usage (e.g., authentication to an app) and is not intended for your first CLI usage.

You can use the -h or --help flag after the command. This will provide you with a list of possible commands and their usage instructions (available options, arguments, …). The basic command is ovhai -h. Additionally, you can refer to the ovhai CLI cheat sheet for a comprehensive overview of available commands.

Although these products are all dedicated to AI, giving you access to powerful resources and the ability to interact with your data, they each have a specific use:

Notebooks are files that contain both computer code (e.g., Python) and rich text elements (paragraphs, equations, figures, links, etc.). This is why Notebooks are mainly used for data exploration, studies, and comparing solutions. OVHcloud AI Notebooks provides you with managed Jupyter or VSCode notebooks, which come with pre-installed libraries. This way, you do not have to worry about your environment, which will save you some time.

AI Training is dedicated to model training. Once your environment has been configured, you can launch a model training session in just a few clicks. As soon as the training is complete, the job will stop automatically. So, it is the perfect tool for training your models without having to worry about the bill.

AI Deploy allows you to deploy your AI applications and models with high availability. If you're looking for a tool to make inferences, this is the product you are looking for.

For more information on these products, please read the comparative guide for AI Solutions.

Currently, the available hardware for AI Tools is:

• CPU: CPU vCores
• GPU: NVIDIA L4

New flavors will soon be available, such as NVIDIA L40S.

You can get the list of available flavors using the ovhai CLI and the following command: ovhai capabilities flavor list.

New flavors will soon be available, such as NVIDIA H100 and also L4 and L40S.

Each Public Cloud project grants a customer, by default, a maximum number of GPUs used simultaneously. Reach out to our support if you need to increase this limitation.

For your information, the current limits are:

• CPU: 12 per notebook / job / app
• GPU: 4 for L4

Each tool provides local storage. However, it is limited and not the recommended way to handle data. Indeed, if you delete your notebook, job, or app, all files contained in this local storage will be lost. A better practice is to attach Object Storage containers.

Read this OVHcloud documentation on data for more information. You can also check the S3™ * compatible compliance documentation if you are interested in S3™ compatible buckets.

You can upload any file or directory to the OVHcloud Object Storage without any format constraints.

Depending on how you build your container, make sure that the mapping between your data (/workspace/mybucket, for instance) does not already exist within your environment.

Within AI Tools, code and users have no root/sudo privileges. To have access to your files, make sure to mount your data object at a location available for non-root users. For preset images provided by OVHcloud, it is recommended that the destination path be of the form /workspace/<your-path> to avoid such errors.

However, jobs in AI Training and apps in AI Deploy are basically Docker containers, so a practical understanding of Docker is required to fully benefit from the service. If you plan to use custom images, you can do the same by adding RUN chown -R 42420:42420 /workspace to your Dockerfile.

While AI Notebooks does not offer this possibility, it is possible with AI Training & AI Deploy. Your image needs to be on the shared registry, a private managed registry, DockerHub, or GitHub packages. You will find more information here. Keep in mind that your image must respect the Linux/AMD64 architecture.

Managed private registry is currently available in Vint Hill (US-EAST-VA region).

Essential information about building a custom Docker image can be found in the Build and use a custom Docker image guide. Advanced information can be found in the official Docker documentation.

For security purposes, we impersonate the Linux default user, which is ovh and group ovh with IDs 42420:42420.

Building a docker with a directory associated with this user and group should help you.

You could use the following command to copy a local folder with the ovh:ovh rights.

COPY --chown=42420:42420 my_local_folder my_folder_embeded_in_image

Start with retrieving the task ID with this command: ovhai <task> list. Make sure to replace <task> with either notebook, job, or app.

Once you have your task ID, simply retrieve its information with ovhai <task> get <task-id>.

You will see, among other information, the task's state and if the task has run its exit code.

There is also the stateInfo in which you can evaluate the error message, i.e., whether a command failed or the Docker image was not found.

For more information about the failure of an AI task, you should consult the Info Url of your task, or its logs, in the case of a job or an app: ovhai <task> logs <task-id> which is very useful to see your prints and error messages.

Your AI Task might be in this state for one of two reasons:

• You are using an external registry, and the image is taking longer to pull. Potential resolution: wait a bit longer for the AI Tool to pull the external image or recompile the image on an OVHcloud Managed Private Registry.
• The AI Tool is waiting for resources to be available. Potential resolution: try to launch the job with fewer resources or wait for resources to be available.

Make sure your AI Task is in a RUNNING state and that your UI is exposed either on the default port or that you specified the correct port in your URL (see public ports).

Only the HTTP layer is accessible and checks that your UI is correctly bound to the network interfaces (e.g., with tensorboard user the --bind-all flag).

When launching an AI Task through the control panel, make sure to select Public access during the Privacy Settings step for AI Notebooks and Training or the Configure your app step for AI Deploy.

If you use the ovhai CLI, make sure to use the --unsecure-http attribute.

This way, your AI Tool HTTP link will not require authentication to be accessed.

• Public networking can be used for all the AI Tools.
• Private networking (OVHcloud vRack) is not supported.

Each notebook has a public URL. By default, this URL accesses the port 8080 of the notebook. The default port cannot be changed.

However, you can access other ports by appending them to the URL. For example, the notebook URL (starting with the notebook's ID, filled with 0 here) for accessing the 8501 port will be https://00000000-0000-0000-0000-000000000000-8501.job.us-east-va.ai.cloud.ovh.us/

Concerning jobs and apps, you can map them to only one port. Each of these has a public URL. By default, this URL accesses the port 8080 of your job or app. This default port can be configured when you submit your job or your app.

As for AI Notebooks, you can also access other ports by appending them to the URL. For example, your default Job URL, starting with the job's ID and which accesses the default port, is https://00000000-0000-0000-0000-000000000000.job.us-east-va.ai.cloud.ovh.us. If you want to access port 9000, you will have to append the port number to your job's URL after the job's unique ID: https://00000000-0000-0000-0000-000000000000-9000.job.us-east-va.ai.cloud.ovh.us.

The same applies to AI Deploy apps, where another app port can be accessed even after the app has been launched.

Although not available with AI Notebooks, you can use gRPC on your AI Training jobs and AI Deploy apps.

To do this, specify a port when you launch your AI Task with the CLI by using the --grpc-port <GRPC_PORT> attribute. You will then get your gRPC Address in your task info.

Yes, you can use the TensorBoard visualization tool. To find out how, check this tutorial. For AI Training, refer to this documentation.

Yes, you can use the Weights & Biases dashboard inside your notebook. To find out how, check this tutorial. For AI Training, refer to this documentation.