Can't hide GPUs to get_cuda_device_count()

Question

Can't hide GPUs to get_cuda_device_count()

mattiadg opened this issue 2 months ago · comments

Hi,
thank you for this great project!
I'm running on Windows 11 and ctranslate2 v4.2.1. When running in a machine with gpus, even if I set CUDA_VISIBLE_DEVICES to the empty string, get_cuda_device_count() always returns the number of gpus in the machine.
Is this expected?

Minh-Thuc · Answer 1 · Mon May 13 2024 23:50:57 GMT+0800 (China Standard Time)

CUDA_VISIBLE_DEVICES should affect which GPU will be visible. If you have more than 2 GPUs, try export CUDA_VISIBLE_DEVICES=1 and run a script to see if GPU 1 is used as expected.

Mattia Di Gangi · Answer 2 · Tue May 14 2024 00:35:29 GMT+0800 (China Standard Time)

I have one gpu only, and when I set CUDA_VISIBLE_DEVICES="" the count is still 1. Il Lun 13 Mag 2024, 17:51 Minh-Thuc ***@***.***> ha scritto:

…

CUDA_VISIBLE_DEVICES should affect which GPU will be visible. If you have more than 2 GPUs, try export CUDA_VISIBLE_DEVICES=1 and run a script to see if GPU 1 is used as expected. — Reply to this email directly, view it on GitHub <#1693 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AA7LDISRBLZNAZFGACHKIQLZCDOPNAVCNFSM6AAAAABHLYM5VCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCMBYGA3TOMZQG4> . You are receiving this because you authored the thread.Message ID: ***@***.***>

Minh-Thuc · Answer 3 · Tue May 14 2024 17:14:07 GMT+0800 (China Standard Time)

I just tested it, if I set CUDA_VISIBLE_DEVICES="" and get number of gpu device with ctranslate2::get_gpu_count() , I got the correct value.

Mattia Di Gangi · Answer 4 · Tue May 14 2024 18:57:24 GMT+0800 (China Standard Time)

did you test on Windows? Il Mar 14 Mag 2024, 11:14 Minh-Thuc ***@***.***> ha scritto:

…

I just tested it, if I set CUDA_VISIBLE_DEVICES="" and get number of gpu device with ctranslate2::get_gpu_count() , I got the correct value. — Reply to this email directly, view it on GitHub <#1693 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AA7LDIS3U7DF375G3TSFJITZCHIXLAVCNFSM6AAAAABHLYM5VCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCMBZGY4DQNRXGQ> . You are receiving this because you authored the thread.Message ID: ***@***.***>

Minh-Thuc · Answer 5 · Tue May 14 2024 20:53:05 GMT+0800 (China Standard Time)

I tested on Linux. Make sure you set correctly the environment variable. You can test directly with CUDA without using ct2 before.