Blender 2.93.5: Node not using GPUs even with correct Cuda version and devices found

[ghost]

I built a new image based on this repo's Dockerfile.cuda11 and https://github.com/zocker-160/blender-docker/blob/master/Dockerfile.cuda11 to use Blender 2.93.5 (latest at this moment). I only modified the major/minor version numbers and set the "local" flag to "true".

My cards are recognized on the host and use Cuda 11

The container also reports those devices:

found following CUDA devices:
[<bpy_struct, CyclesDeviceSettings("NVIDIA GeForce RTX 2080 Ti") at 0x7f6ff663a408>, <bpy_struct, CyclesDeviceSettings("NVIDIA GeForce RTX 2080 Ti") at 0x7f6ff663a488>, <bpy_struct, CyclesDeviceSettings("Intel Core i9-9900KF CPU @ 3.60GHz") at 0x7f6ff663a508>]
dodgy, but trying to unregister anyway
using workaround for source CR version...
starting crowdrender server in LOCAL MODE.....................

I am using a docker-compose file to run the service as such:

services:
  blender-crowd:
    build:
      context: .
    network_mode: host
    environment:
      - token=<redacted>
    container_name: blender-crowd
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

When i run a network render on my master machine, I notice that the docker container on the remote node is only using the cpu to render. Is there a way to use the two GPUs instead or additionally?

If yes, how so?

Best,
Val

EDIT: I had forgotten to setup the node to use Cuda, which is now done, but it doesn't help

[ghost]

I tried to only enable devices that contain "NVIDIA" in the name which makes sure that the CPU is not enabled for Cuda rendering.
Edited activate_gpu.py like so:

for device in bpy.context.preferences.addons['cycles'].preferences.devices:
        if 'nvidia' in device.name.lower():
                print('Using ' + device.name)
                device.use = True

Output is as expected but the GPU(s) still don't get used.

I reverted to activate all 3 devices. The render DOES happen very slowly on the CPU (maybe using Cuda?) if I enable it both in python file and in UI.

GPUs get a blip of utilization for 1 second at the very start then go quiet.