Jupyter with GPUs

Warning

Certain projects have funded hardware for Jupyter with GPUs. The resources are available to all Triton users, with priority given to the project members. Others can attempt to use in a preemptible queue (the jobs are killed with no warning if a higher-priority user comes).

We are still tuning the parameters (run time, resources available, etc.) to balance usefulness vs resource wastage. There is no service guarantee. Let us know what is useful or not working.

The normal OnDemand Jupyter does not include GPUs, because they are very expensive and Jupyter interactive work by its nature has lots of idling. However, some projects have ordered GPUs specifically for interactive work. The GPUs that have been dedicated for interactive Jupyter work are divided into separate virtual GPUs with less GPU-memory.

Expected use case

Remember, Jupyter+GPUs are designed to be used for testing and development, not production runs or real computation. The GPU memory is limited, so you can test code but probably not even run moderately-sized models. This is because any resources allocated to a Jupyter job are mostly idle.

You should plan (from the beginning) how you will transition to batch jobs for your main computations. For example, write and verify code in Jupyter with tiny data, then from the command line submit the code to run in the batch queue with much more resources:

$ sbatch --gpus=1 --wrap 'jupyter nbconvert --to notebook --execute --allow-errors mynotebook.ipynb --output mynotebook.$(date -Iseconds).ipynb'

How it works

  • Use the normal OnDemand Jupyter app, https://ondemand.triton.aalto.fi, as described in Jupyter on Triton.

  • Select one of the interactive partitions (see below)

  • Your Jupyter session will start. Note it has shorter timeouts that other Jupyter sessions, to prevent inefficiency. Once you have resources, don’t forget to use them.

  • When you are done with using the resources, remember to stop the session via File > Shut Down.

  • There is no service guarantee, resources may be stopped or adjusted anytime without warning. Save often.

Name

Who has access

Resources

Ellis H200 GPU

ELLIS project staff (ellis unix group). Still being formed, contact ASC for access for now.

8 H200 GPUs split into a total of 56 vGPUs with 18G mem each.

General H200 GPU

Anyone, but sessions can be stopped without warning if a higher priority user comes and needs the resources.

Same as above

Time limits and other parameters are visible in OnDemand (and not copied here since they may change).

Contact

Contact ASC/Science-IT us via the normal means, or the people in the table above for access to the resources.