nodes show only cpu and cuda:0, missing cuda:1

[Mark-883]

Hi,
i have 2x gpus, 3090 and 4090
Comfyui only detect one of them at time, if i start it with command --cuda-device 0 or 1 it detect 3090 or 4090 but only one.

multigpu nodes always show me only cpu and cuda:0 options, im unable to get cuda:0 and 1 both at the same time, probably because comfyui detect only one gpu

any help?

[951336ggg]

Hi, i have 2x gpus, 3090 and 4090 Comfyui only detect one of them at time, if i start it with command --cuda-device 0 or 1 it detect 3090 or 4090 but only one.

multigpu nodes always show me only cpu and cuda:0 options, im unable to get cuda:0 and 1 both at the same time, probably because comfyui detect only one gpu

any help?

the same issue，have you found the solution？

[Mark-883]

the same issue，have you found the solution？

not yet

[shdeng]

It works to me when I set CUDA_VISIBLE_DEVICES=0,1

[Zibbibbo]

set CUDA_VISIBLE_DEVICES=0,1

i've already tried this but doesn't work for me.

what version of comfyui are you using?

[shdeng]

set CUDA_VISIBLE_DEVICES=0,1

i've already tried this but doesn't work for me.

what version of comfyui are you using?

The latest version. It might be a silly check :D, but you could see the values from echo and then didn't spec --cuda_device once start comfyui?

[Zibbibbo]

The latest version. It might be a silly check :D, but you could see the values from echo and then didn't spec --cuda_device once start comfyui?

if i dont specify --cuda_device it will use by default the 4090, ignoring the 50, and the multigpu node show cpu and cuda:0
Crystool show me both gpus cuda:0 and cuda:1 with relatives separate monitorings, but comfy still ignore one gpu.

btw, my previous question was about, are you using Comfyui Desktop version, or a portable? or something else?
im trying with desktop but this one doesn't accept the command line set cuda visible, and the portable updated to latest nightly build, but still nothing to do for me.

[shdeng]

The latest version. It might be a silly check :D, but you could see the values from echo and then didn't spec --cuda_device once start comfyui?

if i dont specify --cuda_device it will use by default the 4090, ignoring the 50, and the multigpu node show cpu and cuda:0 Crystool show me both gpus cuda:0 and cuda:1 with relatives separate monitorings, but comfy still ignore one gpu.

btw, my previous question was about, are you using Comfyui Desktop version, or a portable? or something else? im trying with desktop but this one doesn't accept the command line set cuda visible, and the portable updated to latest nightly build, but still nothing to do for me.

I am running the cloned source code on Linux, that might be the difference? Here just run "export CUDA_VISIBLE_DEVICES=0,1" in terminal then start comfyui works well.
And crystool use different way to determine the devices, no matter you set --cuda_device or not, the installed gpus should still be detected.

[Mark-883]

With the latest update of Comfyui Desktop App in the "Server-config" settings menu under CUDA leave blank the "Cuda device index to use"

now multigpu nodes show cpu/cuda:0/cuda:1

i don't know if it's for real the latest update or some ot the million try i've done inside the custom nodes folder and settings.json.
previously, until yesterday, also with the cuda index blank did not working

[Zibbibbo]

With the latest update of Comfyui Desktop App in the "Server-config" settings menu under CUDA leave blank the "Cuda device index to use"

now multigpu nodes show cpu/cuda:0/cuda:1

i don't know if it's for real the latest update or some ot the million try i've done inside the custom nodes folder and settings.json. previously, until yesterday, also with the cuda index blank did not working

got it to work too... finally

[falcorr]

main.py:

    if args.cuda_device is not None:
        os.environ['CUDA_VISIBLE_DEVICES'] = str(args.cuda_device)
        os.environ['HIP_VISIBLE_DEVICES'] = str(args.cuda_device)
        logging.info("Set cuda device to: {}".format(args.cuda_device))
    
    os.environ['CUDA_VISIBLE_DEVICES'] = "0,1"
    os.environ['HIP_VISIBLE_DEVICES'] = "0,1"