Installing two DIGITS on one computer

[szm-R]

Hi everyone

I want to install two digits on my system, one to use as CPU-only and the other for GPU .As far as I know to use digits in CPU-only mode one must build nvcaffe in this mode and then link this caffe with Digits and there isn't a way to optionally choose the platform for training. So in order to be able to use both cpu and gpu for training I think I need to install two DIGITS, right? (I know that training on cpu is much slower than gpu, but still I need to do this).

Now the problem is that to start the webapp we enter address localhost:5000, how does this work when there are two digits on system?!

Thanks in advance!

[lukeyeager]

Read these docs:

https://github.com/NVIDIA/DIGITS/blob/digits-5.0/docs/Configuration.md
https://github.com/NVIDIA/DIGITS/blob/digits-5.0/digits/__main__.py

Example usage:

DIGITS_JOBS_DIR=~/digits-jobs-cpu CAFFE_ROOT=~/caffe-cpu ./digits-devserver --port 5001

DIGITS_JOBS_DIR=~/digits-jobs-gpu CAFFE_ROOT=~/caffe-gpu ./digits-devserver --port 5002

[szm-R]

Thank you very much! I will test this.

[szm-R]

Hello again, As I said before I want to use digits in cpu only mode, I have built caffe in cpu only and then when I run digits as you said it still detects gpu (and because the caffe is cpu only training results in error), I searched and there seemed to be this way of running ./digits-devserver --config to choose none as gpu device but when I try to run this it says:
main.py: error: unrecognized arguments: --config
I have done this once (a couple of month before) and it worked but now it gives me this (digits version is 4.1). what should I do?

[lukeyeager]

References to the --config flag are outdated. DIGITS 5.0 includes #1091 which does away with that flag.

It's possible that adding CUDA_VISIBLE_DEVICES= (setting it to the empty string) to the environment of your CPU DIGITS process will help DIGITS recognize that it shouldn't worry about GPUs.

[szm-R]

running with empty CUDA_VISIBLE_DEVICES= solved it, it gives the following error
cudaRuntimeGetVersion() failed with error #38
but then it runs OK on CPU, thanks!