Change Details

The aim of this task is test AMD GPUs on our stack and identify the challenges/blockers (if any) of using them. We want to test the following: - run some open source LLM models - deploy/serve an LLM that has been trained on an Nvidia GPU - Kserve: deploy/serve a model using kserve - Kubeflow: run an example of a training pipeline where the training step/pod uses a GPU. To test this we can install Kubeflow on minikube/kind and make the AMD GPU available.

The aim of this task is to test AMD GPUs on our stack and identify the challenges/blockers (if any) of using them. We want to test the following: - run some open source LLM models - deploy/serve an LLM that has been trained on an Nvidia GPU - Kserve: - deploy/serve a model using kserve - investigate how to share a GPU among multiple models and what the community is doing on this topic. There are two approaches we should explore: - Share a GPU among two or more pods - share GPU among multiple models in one pod. - Kubeflow: run an example of a training pipeline where the training step/pod uses a GPU. To test this we can install Kubeflow on minikube/kind and make the AMD GPU available. Test common frameworks(Pytorch, Tensorflow) with the GPU.

The aim of this task is to test AMD GPUs on our stack and identify the challenges/blockers (if any) of using them. We want to test the following: - run some open source LLM models - deploy/serve an LLM that has been trained on an Nvidia GPU - Kserve: - deploy/serve a model using kserve - investigate how to share a GPU among multiple models and what the community is doing on this topic. There are two approaches we should explore: - Share a GPU among two or more pods - share GPU among multiple models in one pod. - Kubeflow: run an example of a training pipeline where the training step/pod uses a GPU. To test this we can install Kubeflow on minikube/kind and make the AMD GPU available. To test this we can install Kubeflow on minikube/kind and makeTest common frameworks(Pytorch, Tensorflow) with the AMD GPU available.