femtonelson
commented
2 years ago Hello, We forked the project and tweaked the code to achieve this.
- Added the required MIG GPU references below to labextension files "CellMetadataEditorDialog.tsx" and "InlineMetadata.tsx": nvidia.com/mig-1g.10gb : NVIDIA MIG 10GB nvidia.com/mig-2g.20gb : NVIDIA MIG 20GB nvidia.com/mig-3g.40gb : NVIDIA MIG 40GB
- Built and deployed new npm package : https://www.npmjs.com/package/kubeflow-kale-labextension0.7.mig
- Edited the regex variable LIMITS_TAG in the python Kale backend nbprocessor.py file found in "/usr/local/lib/python3.6/dist-packages/kale/processors" or "/opt/conda/...../kale/processors", ... to support new MIG GPU references
- Finally updated the Dockerfile, rebuilt and published this expleorandd/jupyterlab-kale-cpu:9c80c1a0 docker image to Dockerhub
Special thanks to Melissa for her contribution.
Hello, With NVIDIA's Multi-instance GPU feature that allows GPU splicing, new fully qualified names to identify GPU resources have been introduced. In addition to the traditional "nvidia.com/gpu" resource, other GPU resources as shown below are now schedulable. requests.nvidia.com/gpu: "4" requests.nvidia.com/mig-1g.10gb: "4" requests.nvidia.com/mig-2g.20gb: "4" requests.nvidia.com/mig-3g.40gb: "4"
Could you kindly update the Cell Metadata Editor
to support these new resources? Possibly an empty field to specify the GPU resource in Kale would be great, as there are many new possible MIG GPU instance names.
Kubeflow notebooks allow customization via jupyter-web-app configmap.
It would be great to have such customization option in Kale as well.
Thanks for your support. https://docs.nvidia.com/datacenter/tesla/mig-user-guide/index.html