Why I got 3 other processes on GPU 0 when use DDP to train model?

1

5a40b07f981587ffaa0b7728e265c01
5a40b07f981587ffaa0b7728e265c01882×280 87.7 KB

2

I guess you are initializing a additional CUDA contexts on the default device (most likely from the other processes). Check if you have any CUDA operations working on the default device instead of the one used in the DDP launch.