[Core] CPU count in available_resources() does not match CPU count from psutil

[ericbhanson]

What happened + What you expected to happen

I'm running Ray Tune on a Paperspace machine that has 8 CPUs, but I noticed from the Tune dashboard that there are only 5 CPUs available. If I run psutil.cpu_count(logical=False) or psutil.cpu_count(logical=True) I get a CPU count of 8 as expected, but if I run ray.init() and then ray.available_resources(), I get a CPU count of 5.

Versions / Dependencies

Python: 3.9.16 Ray: 2.40

Reproduction script

import psutil import ray

print(psutil.cpu_count(logical=False)) print(psutil.cpu_count(logical=True))

ray.available_resources()

Issue Severity

Medium: It is a significant difficulty but I can work around it.

[scv119]

@ericbhanson do you have anything else running when you run this script?

[ericbhanson]

@scv119 It’s running inside of a Jupyter notebook. Otherwise I’m not aware of anything else running.

[ericbhanson]

@scv119 I just tried running it within a terminal inside of Paperspace's interface and I'm getting the same CPU count.

[ericbhanson]

@scv119 there's also this message when Ray initializes that might be helpful: Ray currently does not support initializing Raywith fractional cpus. Your num_cpus will be truncated from 5.985 to 5.

[ericbhanson]

Also I want to note that I'm now running on Ray 2.5 instead of 2.4.

[rickyyx]

Thanks @ericbhanson ! I will try repro on my end.

[ericbhanson]

Not sure why, but a more recent release seems to have fixed the issue, as I'm no longer seeing. Closing...