Hello there !
We are really glad to announce the launch of OVHcloud AI Training as a paid beta, fully integrated in our public cloud openstack ecosystem.
You may know that data-related projects are booming, and we want to provide you all the required tools to build you own data and AI dreamed platform.
Since 2 years we added new services (Apache Hadoop big data clusters, Data processing powered by Apache Spark, ML Serving to deploy models in production, …).
Now, with AI Training, we simplify the life of data scientists and data engineers for neural network trainings over GPUs.
You just have to push your code in a docker and your data on Object Storage and the platform takes care of the rest (data sync, workload deployment, user management, ...).
If you don't know docker, it doesn't matter, you can use our prebuilt images with your favorite framework (HuggingFace, Pytorch, Tensorflow, MXNet, Fast.ai, ...) with JupyterLab notebooks and VScode.
All of that with as always, a simple and nice pricing : 1.75€ /hour /GPU NVIDIA V100s (more flavors to come…)
What does it solve ? lot of things ! easiness to scale, orchestration, no more infrastructure to manage, cost control, … focus on code 😊
Don’t hesitate to try it out, to share feedback or questions ! it’s on your Public Cloud control panel (AI training in left menu) or via CLI
Product page : https://www.ovhcloud.com/en/public-cloud/ai-training/
Documentation : https://docs.ovh.com/gb/en/ai-training/
Roadmap : https://github.com/ovh/public-cloud-roadmap/projects/2
Have a good day !
And OVH API answers: ovh.exceptions.BadParametersError: You provided an input body while none was expected
Indeed, I have some more questions :
1/ Is it possible to create block storage via API
2/ and then to attach/mount it on AI Training instance (in addition to some Object storage ones)
3/ do you have any advice to allow multi-user usage on a single AI Training instance ?
Thanks
2) No, but we can discuss on discord about your need if you want: https://discord.gg/uZpsNAK5aU
3) Yes you have to ways to do it, you can create openstack users (AI Training reader role) or you can create token with the CLI.
Hello @XciD thanks for the answer.
Is it possible to get filtered job according to some fields (I only see status and updatedAfter atm) => for instance, we want to create an AI Training instance for a specific user; is is possible to (filtered) list job of a specific user to master the fingerprint of the result? (instead of listing 100% jobs and filtering on our side)
Is is possible to renew the AI Training instance timeout? Or is there any possibility to postpone the shutdown ?
Hello, some other questions: we tried to mount the same 'Object Storage' (for same user, RW permission, and same prefix) on several AI Training instances; but the files/data created on them seem not updated in real time ?
During our first tests, after the end/death of AI Training instances, there was nothing on the share; but it seems OK now (files are available after new AI Training instance creation with same share).
2/ we tried it simultaneously on 2 running AI Training instances, and nothing is visible on the other side (file1 created on instance1 is not shown on instance2, and vice versa); may be another way to see the same issue.
=> after creation of new AI Training instances, with same share, then we can see files created from both the previous instances.
How could we quickly fix that please (to get a real-time workspace/share space)?
Hello @bertrand-benoit_gitlab, right now it is indeed only possible to filter jobs based on status and update date. We will add the ability to filter based on users in our backlog.
It is not possible at the moment to update the timeout of a job, the feature is planned but was not prioritised so far, we'll move it up.
About the sharing, upon mounting an Object Storage container on a job, you need to add the cache option to have the same mount on every jobs. If the cache option is not specified it is actually independent copies of the same Object Storage container that are mounted on each job.
ovhai CLI you can do so by resubmitting from an existing job and overriding the ressources:ovhai job run --from <job-id> --cpu 2
Hi,
I'm testing the Notebook AI beta. i'm not sure if this is the right place to ask questions about this service. Please tell me if there is a more appropriate chat.
Question: I'm trying to connect to the Jupyter Notebook with my credentials.
The doc states:
Any person who is in the same Public Cloud project as you will also have access to this notebook using their own credentials
My NIC-handle is allowed in the Public Cloud / Project Management / Contacts and Rights section with Rread/Write access. Even though, the Jupyter sign in fails.
Invalid user or password
If I create a user in Public Cloud / Project Management / Users & Roles (with AI Training Operator and Object Storage operator), I can sign in onto the Jupyter Notebook.
What am I doing wrong?
In the meantime it is possible to AI Training to deploy your API while leveraging GPUs, you can find an example of how to do so here: https://github.com/christophe-rannou/sample_vision_api_yolov4 and I can provide you additional help if needed.
Do not hesitate to join us on the Discord server: https://discord.com/invite/vXVurFfwe9 (on which we are more active lately)
