Kubeflow listed as available AArch64 addon, but missing enable script
See original GitHub issuePer #1086, and my (limited) understanding from Kubeflow discussion elsewhere, I gathered that KF probably wouldn’t be available on AArch64 MicroK8s. I was surprised then when microk8s enable --help
showed Kubeflow on the list, and also when I double-checked the supported architectures in addon-lists.yaml
.
However, when I tried running microk8s.enable kubeflow
, I got Nothing to do for kubeflow
, which I believe I’ve traced back to the fact that ${SNAP}/actions/enable.kubeflow.sh
doesn’t exist.
Is this expected behavior? I mostly just wanted to report this since I found it odd that KF is ostensibly listed as an available add-on, but the script to enable doesn’t seem to be shipped with the snap I installed. I suspect I may just be misunderstanding something here too.
I believe I’m using a fresh installation of the latest stable version as of today, but here’s the inspect archive for good measure. Thanks!
Issue Analytics
- State:
- Created 3 years ago
- Comments:7 (2 by maintainers)
Top GitHub Comments
@matthiaslau If I remember correctly, I did tests on that (I guess by saying distributed training you are mentioning deploy kubeflow on a multi-node k8s cluster and run training job there).
What I did was running the object detection example using self-compiled tensorflow 1.10 GPU version container and deployed everything with v0.5.1 kubeflow on my cluster (an arm64 server as master and Jetson TX2 as node). I did not have problem with TFJob itself, the training can also be started, but after (I think) several steps it will warn me about the OOM error, which to me is a little bit weird…
I am pretty sure that I used the GPU as I think the GPU version tensorflow will automatically assign tasks to the GPU, and I could see the GPU0 in the log…
Thank you for the answer! 😃 I am currently working with a cluster of NVIDIA Jetson Nanos (GPU: Maxwell with 128 Cores, CPU: Quad-Core-ARM A57 with 1,43 GHz). It is also interesting to expand the cluster with Jetson Xavier (GPU: Volta with 512 Cores, CPU: 8 ARM-v8.2-64-Bit-CPU-Cores).
It would be really great if GPU and Kubeflow will be supported in the future as microk8s would be a good fit for this hardware. But I see the difficult dependencies especially for kubeflow.