Added the possibility to filter by GPU IDs

[PCudrano]

The standard nvidia-smi summary provides an option to display only some of the available GPUs. This feature is very useful in systems with many GPUs.

The standard nvidia-smi command is:
nvidia-smi -i <comma-serparated-gpu-ids>
For example, to display only GPU 1, 2 and 4:
nvidia-smi -i 1,2,4

I added the same exact -i argument, exploiting the nvidia-smi command directly.

Usage:

nvidia-htop.py -i <comma-separated-gpu-ids>

For example, to display only GPU 1, 2 and 4:
nvidia-htop.py -i 1,2,4

If a wrong ID list is passed, the command fails with the original nvidia-smi error message.
For example, asking for an non-existing GPU ID:
nvidia-htop.py -i 42 or nvidia-htop.py -i random,string
Return:
nvidia-smi exited with error code 6: No devices were found

[peci1]

Thanks for the PR.

Could you please also update the readme? And, there is also a description of the CLI options at the beginning of the Python script. Could you update that, too?

Would it be possible to add a test for this CLI flag? I think you can just copy one of the already present test and just add the -i flag. This is to test that the CLI flag is correctly parsed and recognized.

[peci1]

And do you know since when the -i flag is supported by nvidia-smi? Has it been there forever?

[PCudrano]

My pleasure, I've been using the script for some time and figured I'd contribute with my little edit.

I added the usage string in the readme and in the Python script, thanks for pointing it out.

It's worth stressing out that this implementation sends the -i argument directly to the nvidia-smi call by subprocess.run. This means that the -i option would not work if this command is piped to nvidia-smi. It only works when nvidia-htop.py is called directly. Otherwise, this option is not considered.
We might look into filtering out the GPU IDs manually from the full nvidia-smi output, but that'd be kinda inefficient, I don't see much purpose to it if not for backward compatibility. Let me know what you think though.

I looked into testing the changes, but for the reason above I could not add new test cases, as they'd require nvidia-htop.py to be piped to a fake output txt.

I think I've used the native -i option on nvidia-smi for years, but I'm not sure when it was first introduced tbh.

[peci1]

Thanks for pointing out it doesn't work in pipe mode. I've added a warning. I also added the test, but it only tests that the option is accepted, nothing more.

[peci1]

Released in version 1.1.0. Thanks for your contribution!

[PCudrano]

Great! Thank you!