Hmm. I see that when you use

GPUAccelerator

like in:

@task(container_image=image_spec, requests=Resources(cpu="1", mem="2G"), accelerator=GPUAccelerator("nvidia-l4"), limits=Resources(cpu="4", mem="7G", gpu="1"))

It will set

affinity:
    nodeAffinity:
      requiredDuringSchedulingIgnoredDuringExecution:
        nodeSelectorTerms:
        - matchExpressions:
          - key: <http://k8s.amazonaws.com/accelerator|k8s.amazonaws.com/accelerator>
            operator: In
            values:
            - nvidia-l4

in the podspec of the task, (even when you're not on AWS...)

Hey Cornelis, could you collect all the gaps you may find in the GPU Accelerators documentation on an Issue? It's a great feature that I think deserves a more descriptive doc

Honestly, that's it. The tolerations explanation is clear here https://docs.flyte.org/en/latest/user_guide/productionizing/configuring_access_to_gpus.html. Just need to add

gpu-device-node-label

and

gpu-partition-size-node-label

such as here https://github.com/flyteorg/flyte/blob/b6f35add1227c930e9208103235d89fe4b864b8e/charts/flyte-binary/gke-starter.yaml#L82-L83 to make it work fully on other platforms than AWS.