Why .cuda() always allocate GPU:0?

kangkang_li · June 21, 2019, 9:38am

Only b.to(device) can move b to GPU:1?

sai_tharun · June 21, 2019, 9:57am

if u have 4 graphic cards,u can choose which one to use.
By deault it will be cuda:0

kangkang_li · June 21, 2019, 10:25am

I have two,so is there a way to change the default gpu?

sai_tharun · June 21, 2019, 10:31am

you can try this
torch.cuda.set_device(1)

JuanFMontesinos · June 21, 2019, 10:55am

It’s easier to use the flag CUDA_VISIBLE_DEVICES=‘1’ rather than coding using set_device.

If you need that it means you are working with a single gpu.

You can call .cuda(idx) which allocates in the gpu you want if you use several gpus.

Deepali · June 21, 2019, 11:44am

CUDA_VISIBLE_DEVICES=“1” exposes only 1 GPU for the program. However, when torch uses the GPUS it starts ordering from 0. So GPU 1 on system is mapped to GPU 0 in the program.

Dreamer_dev · January 13, 2023, 4:59am

Your advice saved me a lot of time and energy Thanks, man! I used the command ‘set CUDA_VISIBLE_DEVICES=1’ before running the code to specify the GPU.