hi @<1729671499981262848:profile|CooperativeKitten94> did i convince you with my argument ? do you think having runtimeClass configurable is worth it ?

well it doesn't fail. but whatever i set gets ignored

this seems to be confirmed by this documentation None If you have not changed the default runtime on your GPU nodes, you must explicitly request the NVIDIA runtime by setting runtimeClassName: nvidia in the Pod spec:

this is my cmdline: clearml-task --name hla --requirements requirements.txt --project examples --output-uri http://clearml-fileserver:8081 --queue aws-instances --script keras_tensorboard.py

AgitatedDove14 your trick seems to work (i had to change the url to reflect the fact i run on k8s)

but i still think the same should be possible using the Task.init

no i don't think so, i think rather Task.init is only used for running outside of agent

and when i try to use --output-uri i can't pass true because obviously i can't pass a boolean only strings

and also, on the tutorials that do something with task.init, the example always talks about running locally and not in the agent

i can pass any crazy value i want.. it doesn't matter. however, i can use --output_uri= s3://blabla and then at least i get the error that it cannot use that bucket

well it made a difference (the code for the init() is not added anymore) but it still didn't take my output uri

hello, i'm still not able to save clearml models. They are generated and registered okay, but they are not on the fileserver. i now have Task.init(output_uri=True) and i also have --skip-task-init in clearml commandline so that it doesn't overwrite the task.init call

this is the script shown by clearML ui. so the task.init call looks right

the model has this information ... the /tmp seem local URIs suggesting that it doesn't even try to upload them

it seems that whatever i pass to Task.init is ignored

this is the output of the training. it doens't try to upload (note that this is my second try so it already found a model with that name, but on my first try it didn't work either)

i set reuse_last_task_id to false to force creation of a new task in all cases

don't know.. but i see for instance when using clearml-task i can put any (even nonsensical) values in Task.init

task = Task.init(project_name='examples', task_name='moemwap', output_uri=True, reuse_last_task_id=False)

i added --skip-task-init too

and ... clearml-agent takes a --project and a --name argument that are mandatory , so these are never taken from Task.init

it didn''t make a difference

so it seems that it takes output_uri from the clearml commandline but not from the Task.init inside the scripot

this is now in my python script:

its as if the line is not there

i'm still trying to understand why it was needed in our case. i have a nvidia gpu operator with mostly the default values installed on our on prem cluster. i found there is an option CONTAINERD_SET_AS_DEFAULT in the operator, which, when enabled, puts the runtime for all pods. we didn't enable that option, maybe if we had enabled it would have worked.

for comparison: this is when i use --output-uri

it was to test if reuse_last_task_id made any effect (i have the impression it doesn't)

(same for environment variable)