Hi @<1800699527066292224:profile|SucculentKitten7> , I think you're confusing the publish action to deployment. Publishing a model does not deploy it, it simply changes the state of the model to published so it cannot be changed anymore and also publishes the task that created it.

To deploy models you need to either use clearml-serving or the LLM deployment application

Hi @<1529271098653282304:profile|WorriedRabbit94> , you can sign up with a new email

It isn't a bug, you have to add the previews manually through reporting. For example:

ds = Dataset.create(...) ds.add_files(...) ds.get_logger().report_media(...)

What is being reported that isn't auto-logged?

Is this something that the ClearML team could add? Would it be helpful to have a call?

FrustratingShrimp3 , I think you can open a github issue to request the new feature

Is it possible the machines are running out of memory? Do you get this error on the pipeline controller itself? Does this constantly reproduce?

Did you raise a serving engine?

You have a self hosted server, several agents that are running and two users developing on their own machines but the training is done elsewhere?

FierceHamster54 , please try re-launching the the autoscaler, the issue seems to be resolved now

MinuteGiraffe30 , Hi ! 🙂

What if you try to manually create such a folder?

Hi @<1579280543999070208:profile|SourFly7> , this index holds scalars of some experiments. You can reduce it by deleting some experiments. Do you have any other large scalar indices?

Are you running in docker mode? You could maybe use another docker image that has python in it.

Basically the Agent automates the docker run command with everything that you need (this can become rather complex). You can see this in the third line of the console log:
` Executing: ['docker', 'run', '-t', '--gpus', 'all', '-l', 'clearml-worker-id=ip-172-31-28-179:0', '-l', 'clearml-parent-worker-id=ip-172-31-28-179:0', '-e', 'CLEARML_WORKER_ID=ip-172-31-28-179:0', '-e', 'CLEARML_DOCKER_IMAGE=nvidia/cuda:10.2-cudnn7-runtime-ubuntu18.04', '-e', 'CLEARML_TASK_ID=cdbfa9cda5ab4d86b012a87...

Hi UnevenDolphin73 , yes this is correct. Artifacts are outputs so they are all removed during a clone or reset

UnevenDolphin73 , I've encountered a similar issue with s3. I believe it's going to be fixed in the next release 🙂

I think the latest is 1.15.0, can you try updating?

ShakyJellyfish91 , Hi!

If I understand correctly you wish for the agent to take the latest commit in the repo while the task was ran at a previous commit?

I would suggest considering a Scale/Enterprise license. The amount of time you would be wasting in man hours on managing, doing hacks and patches on this and various other capabilities would most likely cost you in the long run much more than a license. Not to mention the dedicated support and other features you would be receiving.

Yeah I think you're right. You're basically looking for a capability to segregate users into groups so they won't have visibility between themselves?

Role based access controls are supported only in Scale/Enterprise versions.

You can edit the mongodb manually (strongly suggest against) to change users of experiments. Besides that, I'm afraid not. Each user would have to create separate credentials for themselves under their own user in the system.

A suggestion I might have is using the 'Description' field to write down the relevant user manually and adding that as a column in your view. The small cogwheel near the top right (next to the refresh button) will give you the option to add that column.

Hope this helps...

HugeArcticwolf77 , what if you specify the exact python binary from conda using CLEARML_AGENT_SKIP_PIP_VENV_INSTALL?

Hi :)

I found a comparison here:
https://clear.ml/blog/stacking-up-against-the-competition/

As far as I am aware, there is an on-prem enterprise solution

ShinyLobster84 , sorry for the delay, had to look into it 🙂

Please try task.get_reported_scalars()

I think you would need to add some 'pre' steps. So you would want to build the package from the repository ( python setup.py bdist_wheel ) and then you can either install it manually via the startup script OR add it as a requirement using the following syntax in requirements file:///srv/pkg/mypackage

You can also use Task.force_requirements_env_freeze to freeze an exact copy of your environment.
None

Hi @<1540867420321746944:profile|DespicableSeaturtle77> , what didn't work? What showed up in the experiment? What was logged in the installed packages?

Hmmmmm do you have a specific usecase in mind? I think pipelines are created only through the SDK but I might be wrong

Hi @<1523701083040387072:profile|UnevenDolphin73> , I think you can play with the auto_connect_frameworks parameter of Task.init()
None

Hi @<1523701295830011904:profile|CluelessFlamingo93> , when running remotely the agent assumes it will be a different machine. I think the best way to solve this is to add utils to your repository and import it from there during code execution.

What do you think?