Thank you for your pull request and welcome to our community. We could not parse the GitHub identity of the following contributors: steve.
This is most likely caused by a git client misconfiguration; please make sure to:

1. check if your git client is configured with an email to sign commits git config --list | grep email
2. If not, set it up using git config --global user.email email@example.com
3. Make sure that the git commit email is configured in your GitHub account settings, see https://github.com/settings/emails

Thank you for your pull request and welcome to our community. We could not parse the GitHub identity of the following contributors: steve.
This is most likely caused by a git client misconfiguration; please make sure to:

1. check if your git client is configured with an email to sign commits git config --list | grep email
2. If not, set it up using git config --global user.email email@example.com
3. Make sure that the git commit email is configured in your GitHub account settings, see https://github.com/settings/emails

recheck

From my perspective, this PR has two issues:

• When enumerating GPUs, low-power ones should come first. That way on Intel Macs we’ll use the iGPU instead of the dGPU.

• On Intel Macs, metal::Device::system_default() actually always returns the dGPU, which means the code for selecting another GPU is never reached, so this PR will instroduce a regression.

@JunkuiZhang thank you, I have fixed the issues. is there a way for people who have the issue to try a build?

I think your changes make things a bit more complicated than they need to be. If I understand correctly, the issue on your machine is that devices.pop() returns None, which leads to an error being logged and then an immediate exit. A better approach, in my opinion, would be to log the error but not exit right away. Instead, fall back to metal::Device::system_default().

thank you. updated.

macOS: Enumerate GPUs first; prefer low-power non-removable; fall back to system default
^

Release Notes:

- Added/Fixed/Improved ...

Release Notes:

- N/A

Have feedback on this plugin? Let's hear it!

Generated by 🚫 dangerJS against 38bf822

recheck

I pushed a few changes to your PR, could you check if everything still runs fine on your machine?

thank you! yes it still works for me on MacBook Pro M1 Max. I've never been able to reproduce the original issue. your change is good, but sorting with (is_removable(), is_low_power()) then pop() still risks picking a removable eGPU when present.

proposed change:
let mut devices = metal::Device::all();
let device = if devices.is_empty() {
log::error!("unable to enumerate Metal devices; attempting to use system default device");
metal::Device::system_default().unwrap_or_else(|| {
log::error!("unable to access a compatible graphics device");
std::process::exit(1);
})
} else {
// Prefer iGPU by default; avoid eGPU unless it’s the only device
devices.sort_by_key(|d| (d.is_removable(), !d.is_low_power()));
devices.remove(0)
};

Why:
Prefer non-removable, low-power (iGPU) on Intel for stability/power.
Avoid eGPU by default; use dGPU/iGPU first.
If enumeration is empty (seen in some environments), log and fall back to system_default().
Exit only if both enumeration and system_default() fail.

The change you suggested seems logically the same as mine, the only difference is that you’re using !d.is_low_power() while I’m using d.is_low_power(). From my testing, d.is_low_power() seems to be the correct one. But maybe I misunderstood your point, happy to adjust if that’s the case.

I cleaned it up a bit more, I hope that's okay. My testing on MacBook Pro M1 Max always works, I was not able to reproduce the original issue.

Prefer non-removable, low-power GPU (iGPU) by ordering with min_by_key(|d| (d.is_removable(), !d.is_low_power())), yielding iGPU → dGPU → eGPU.
If enumeration returns empty (as seen in #37689), log and fall back to metal::Device::system_default(); exit only if both paths fail.
Fixes the prior sort+pop behavior, which could select an eGPU/dGPU unintentionally and is harder to reason about.
Avoids Intel regression where system_default() often picks the dGPU; Apple Silicon (single GPU) behavior is unchanged.