when -e DEVICES=all or -e DEVICES=0,1.. env args are set, a corresponding number of processes is spawned, one per GPU. Each process handles copying and moving the model to their respective device, allowing even workload split during inference (increase Number of images requested).
when
-e DEVICES=all
or-e DEVICES=0,1..
env args are set, a corresponding number of processes is spawned, one per GPU. Each process handles copying and moving the model to their respective device, allowing even workload split during inference (increaseNumber of images
requested).