Smallest possible Cloud Run GPU check: an HTTP server that says hello and prints nvidia-smi.
Deploying this once also auto-grants the L4 GPU quota in the region.
gcloud run deploy hello-gpu \
--source . \
--region us-central1 \
--gpu 1 --gpu-type nvidia-l4 --no-gpu-zonal-redundancy \
--cpu 4 --memory 16Gi \
--max-instances 1 \
--allow-unauthenticated
Then curl the service URL and you should see the L4 in the nvidia-smi table. Scale-to-zero is
the default, so an idle service costs nothing.