GPUs follow the engine's own CUDA_VISIBLE_DEVICES

Its entries are the cards present, and a local child's allocation is
named through them, so under =2,3 card 0 is device 2. Remote workers
keep mapping against their own mask.

docs/code/cli.md

@@ -65,8 +65,9 @@ that instance's running engine; plain mode refuses to start a duplicate.
 | `--worker-max-rss MB` | `0` (disabled) | Retire a worker above this resident memory after its call |
 
 Resource defaults can also come from the corresponding `FLOW_*` settings;
-see [Configuration](../reference/configuration.md). GPU detection uses
-NVIDIA device nodes. Set `--gpus` when your hardware is not detected.
+see [Configuration](../reference/configuration.md). GPU detection counts the
+entries of `CUDA_VISIBLE_DEVICES` when it is set, otherwise NVIDIA device
+nodes. Set `--gpus` when your hardware is not detected.
 
 Use local disk for the data directory on a cluster: SQLite's write-ahead log
 does not support NFS.