and just making sure - by pipeline we're talking about the ClearML pipelines, correct?
https://clear.ml/docs/latest/docs/references/sdk/automation_controller_pipelinecontroller

nice! Thank you

I'm sorry. I think I wrote something wrong. I'll elaborate:
The SDK detects all the packages that are used during the run - The Agent will install a venv with those packages.
I think there is also an option to specify a requirements file directly in the agent.

Is there a reason you want to install packages from a requirements file instead of just using the automatic detection + agent?

Hi AverageRabbit65 , can you elaborate on what you're trying to do?
ClearML-Agent will automatically create a venv and install everything

Have you run experiments with the SDK? i.e added Task.init()

now the question how to make the production stuff to work

Yep Correct

or how can I add some libraries I'd like it to pip install in the new environment

Task.add_requirements

based on what reqeuirment.txt manifest?

The one sitting in the repository

I have a simple pipline that works

consider this case:

locally and remotely

PIP Mode

By default, ClearML Agent works in PIP Mode, in which it uses

as the package manager. When ClearML runs, it will create a virtual environment (or reuse an existing one, see

). Task dependencies (Python packages) will be installed in the virtual environment.

how does it know what are the dependencies of a task?

we have a diff name for this file in the repo

I think this is what you're looking for - https://clear.ml/docs/latest/docs/references/sdk/task#taskforce_requirements_env_freeze

how this is possible?

if a==True: import torch else: import tensorflow

the question how does ClearML know to create env and what files does it copy to the task
/homes/yossi/.clearml/venvs-builds/3.7/

the question how does ClearML know to create env and what files does it copy to the task

Either automatically detecting the packages in requirements.txt OR using the packages listed in the task itself

I think I am missing something

from clearml import PipelineController

the only way any computer could figure this out is by running it

And the pipeline runs with agents or locally?

if this was possible we wouldn't need pip

it's an undecidable problem

so it looks for requirement.txt?