Cannot set special GPU usage in pytorch 0.3.1?

1

I am run my code on server that does not allow to upgrade to pytorch 0.4.1. It is still using version of 0.3.1. It has multiple GPU (4 GPUs)

|   0  GeForce GTX TIT...  Off  | 0000:05:00.0     Off |                  N/A |
| 72%   86C    P2   209W / 250W |  12065MiB / 12204MiB |    100%      Default |
+-------------------------------+----------------------+----------------------+
|   1  GeForce GTX TIT...  Off  | 0000:06:00.0     Off |                  N/A |
| 80%   87C    P2   229W / 250W |  11714MiB / 12206MiB |    100%      Default |
+-------------------------------+----------------------+----------------------+
|   2  GeForce GTX TIT...  Off  | 0000:09:00.0     Off |                  N/A |
| 48%   82C    P2   193W / 250W |  11714MiB / 12206MiB |     90%      Default |
+-------------------------------+----------------------+----------------------+
|   3  GeForce GTX TIT...  Off  | 0000:0A:00.0     Off |                  N/A |
| 22%   26C    P8    16W / 250W |      2MiB / 12206MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+

I want to use the GPU 3 only because it is free now. I have set up in python code as

os.environ['CUDA_VISIBLE_DEVICES']=3
torch.cuda.set_device(3)
model.cuda()

I have to print successful the current GPU device as

torch.cuda.device_count()      =', 4L
torch.cuda.current_device()    =', 3L

However, it still use GPU 0 as defalut. How should I correct the above code to use GPU #3?

2

Make sure to set the visible devices before any other imports.
The easiest way would be to run your whole script using:

CUDA_VISIBLE_DEVICES=3 python script.py args

Also note, that setting the visible devices to a particular device masks all other devices.
That means you should only see the one device now with id 0 in your python script.
torch.cuda.set_device(3) shouldn’t work therefore.

3

Thanks. My problem has solved because of PID different