[davidtai]

Proposed changes

GPU.deviceInfo() calls MTLCreateSystemDefaultDevice() on every call. In a command-line process (not an app bundle), the first such call opens a window server session. That step can stall for many seconds.

A sample of a stalled process shows this stack:

GPU.deviceInfo()                      (Source/MLX/GPU+Metal.swift)
 MTLCreateSystemDefaultDevice
  SLSCreateDefaultMetalDevice
   SLSMainDisplayID
    SLSInitialize                      (dispatch_once)
     CGSDisplayInitialize / _CGSGetEventShmem
     TCCAccessPreflight

In the unified log, the process logs TCCAccessRequest for kTCCServiceListenEvent and a com.apple.SkyLight session line, and then nothing until the call returns. On an M5 Max, over 25 runs of the same command-line process, the gap from that SkyLight line to the next event was 0.35 to 0.89 s in 23 runs, 7.1 s once and 30.1 s once. In our case the call happened inside a timed step, so the 30 s stall failed the run.

MLX itself does not use the window server to pick its device. metal::load_device (mlx/backend/metal/device.cpp) takes the first device of MTL::CopyAllDevices() and uses MTL::CreateSystemDefaultDevice() only when that list is empty.

This PR:

• selects the device the same way as metal::load_device: MTLCopyAllDevices().first, else MTLCreateSystemDefaultDevice(). MTLCopyAllDevices is macOS-only, so other platforms keep MTLCreateSystemDefaultDevice().
• computes the value once per process (private static let). The device properties it reads (architecture, maxBufferLength, recommendedMaxWorkingSetSize, HW_MEMSIZE) do not change while the process runs.
• keeps the public API and the returned fields the same.

Testing

• Added MemoryTests.testDeviceInfo: on a GPU device, deviceInfo() reports a real architecture, a nonzero buffer size and working set size, and a second call returns the same values.
• xcodebuild build-for-testing -scheme mlx-swift-Package -destination 'platform=macOS', then xctest for CmlxTests (1 passed), MLXTests (385 executed, 2 skipped, 0 failures) and MLXIntegrationTests (669 passed).
• xcodebuild build -scheme mlx-swift-Package -destination 'generic/platform=iOS' succeeds.
• A small command-line probe on an M5 Mac: MTLCopyAllDevices().first and MTLCreateSystemDefaultDevice() return the same device, architecture, maxBufferLength and recommendedMaxWorkingSetSize. The unified log for the MTLCopyAllDevices path has no TCC or SkyLight lines. The MTLCreateSystemDefaultDevice path logs TCCAccessRequest(kTCCServiceListenEvent) and a SkyLight session.

Checklist

Put an x in the boxes that apply.

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)

🤖 Generated with Claude Code

[davidkoski]

You might also want to cover macCatalyst here too -- the same calls are available but it isn't os(macOS):

#if os(macOS) || targetEnvironment(macCatalyst)

Though if the primary concern is for command line tools, I am not sure you can create one that is macCatalyst.

[davidkoski]

Change looks good, but see comment on macCatalyst -- change that or let me know if you think it isn't needed.

Thank you!

[davidkoski]

I am going to merge as-is -- the macCatalyst change can be made as another PR if needed, but given how those binaries work it may not be a useful case.