Toggle sidebar Toggle sidebar
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.

macOS Golden Gate 27.0 Beta 3 - What's New?

Status
The first post of this thread is a WikiPost and can be edited by anyone with the appropiate permissions. Your edits will be public.
I feel like post-Neural Engine hardware -- basically the M5 Pro/Max with the full dose, M4 with partial -- is going to have a huge advantage over what preceded it. The decision to pick bones over quad-core versus dual-core neural engines, basically matrix math co-processors, in the Watch tends to point to this scenario too, as does this sense of needing 12GB or in other words what we'd have previously seen as Android levels of memory in iPads and iPhones for the full benefit.
 
Why are all of the system icons so inconsistent when it comes to size (I have used the new battery icon as the baseline, as it looks smaller than the rest to me – which indeed it is)?

Screenshot 2026-07-13 at 15.19.20.png
 
Stem splitting not fixed. This was working in beta 2 and there has not been any updates to Logic Pro since, so not sure why people are seeing this in Tahoe
 
Bless you Fable 5:

This crash isn't your fault, and it's not really Logic Pro's fault either — it's a bug in Apple's GPU machine-learning framework on the macOS beta you're running.

Reading the crash​

Three things at the top frame the whole report:

  • Exception: EXC_BAD_ACCESS ... KERN_INVALID_ADDRESS at 0x0000000000000000 — the classic null-pointer dereference. Some code tried to read from address 0.
  • Triggered by Thread 31, a background dispatch queue — not the main UI thread, not the audio engine.
  • OS Version: macOS 27.0 (26A5378n) — the trailing letter on the build number means this is a beta/seed build of macOS 27.

What Thread 31 was doing​

Reading the crashed thread's stack bottom-up tells you the exact chain of events:

  1. MADSP::AudioSourceSeparationServiceBase:😛rocess / ExtractStemsInternal — Logic's Stem Splitterservice kicked off, exactly matching what you were doing.
  2. It hands the audio to CoreML (MLModel predictionFromFeatures) — stem splitting is an ML model.
  3. CoreML routes the model to the GPU via Espresso → MetalPerformanceShadersGraph (MPSGraph).
  4. MPSGraph is compiling the model — the frames full of mlir::... are Apple's ML compiler optimizing the graph. The very top frames are the smoking gun:
FoldMultiplyIntoSDPAScale::matchAndRewrite(ScaledDotProductAttentionOp, ...)<br>mlir::Operation::create(...)<br>mlir::detail::OperandStorage::OperandStorage(...) ← crashed here, reading NULL<br>
An optimization pass called FoldMultiplyIntoSDPAScale (it merges a multiply into the "scaled dot-product attention" step of a transformer model) hit a null pointer while rewriting the graph. That code lives in /System/Library/Frameworks/MetalPerformanceShadersGraph.framework — an OS framework, version 7.0.75.2 from the beta, not part of Logic Pro.

★ Insight ─────────────────────────────────────

  • Stem Splitter is a transformer-based (attention-using) neural network. Before running it on the GPU, macOS JIT-compiles the model through an MLIR pipeline — the same compiler infrastructure LLVM uses — and crashes in an optimizer pass, before your audio is even fully processed.
  • The "Translation fault, byte read at 0x0" plus far: 0x0 in the register dump confirms it read through a null pointer rather than, say, corrupted memory — typical of a compiler-pass bug on a graph shape it didn't expect.
  • Because the pass is new-ish (SDPA folding is a recent MPSGraph optimization), this smells like a fresh regression in the macOS 27 beta.
    ─────────────────────────────────────────────────

What you can do​

  1. File it with Apple via Feedback Assistant — attach this .ips file. Beta-OS crashes in system frameworks are exactly what the beta program is for, and this one has a clean, symbolicated stack pointing at their code.
  2. Wait for the next macOS 27 beta / Logic update — the fix has to come from Apple either way (an MPSGraph fix in the OS, or Logic/CoreML sidestepping the bad pass).
  3. Workaround in the meantime: run Stem Splitter on a machine with a release macOS (26.x), or split the stems with a third-party tool (e.g. Demucs-based apps) until the beta is fixed. There's no Logic preference to force the stem model onto CPU, unfortunately.
  4. Rolling back to the stable macOS release is the guaranteed fix if stem splitting is blocking your work.
Nothing in the report implicates your project file, plugins, or hardware — the MIDI control-surface plugins and Audio Units in the binary list were all idle, parked threads.
 
  • Like
Reactions: Michaelhuisman
mari3824 said:
I've been testing the beta for a couple of days and overall it feels pretty stable on my M2 MacBook. One thing I always check after a major update is whether my everyday apps still work correctly, especially those for screen mirroring and media streaming.

So far, apps from Electronic Team have been working fine for me. I occasionally use their free TV Mirror app, and DoCast has been reliable when casting content from my iPhone to a Smart TV. If anyone is looking for compatible screen mirroring tools after updating, this page may be useful

Has anyone noticed any compatibility issues with third-party apps on Beta 3?
Click to expand...
SyncTwoFolders doesn't work, crashes soon after launch, and any changes made are not saved. Developer has been notified, but he can do nothing until his development software is updated to include GG.
I assume there are other apps that don't work well too.
Check the "Working" thread for more info.
 
ErikGrim said:
Bless you Fable 5:

This crash isn't your fault, and it's not really Logic Pro's fault either — it's a bug in Apple's GPU machine-learning framework on the macOS beta you're running.

Reading the crash​

Three things at the top frame the whole report:

  • Exception: EXC_BAD_ACCESS ... KERN_INVALID_ADDRESS at 0x0000000000000000 — the classic null-pointer dereference. Some code tried to read from address 0.
  • Triggered by Thread 31, a background dispatch queue — not the main UI thread, not the audio engine.
  • OS Version: macOS 27.0 (26A5378n) — the trailing letter on the build number means this is a beta/seed build of macOS 27.

What Thread 31 was doing​

Reading the crashed thread's stack bottom-up tells you the exact chain of events:

  1. MADSP::AudioSourceSeparationServiceBase:😛rocess / ExtractStemsInternal — Logic's Stem Splitterservice kicked off, exactly matching what you were doing.
  2. It hands the audio to CoreML (MLModel predictionFromFeatures) — stem splitting is an ML model.
  3. CoreML routes the model to the GPU via Espresso → MetalPerformanceShadersGraph (MPSGraph).
  4. MPSGraph is compiling the model — the frames full of mlir::... are Apple's ML compiler optimizing the graph. The very top frames are the smoking gun:
FoldMultiplyIntoSDPAScale::matchAndRewrite(ScaledDotProductAttentionOp, ...)<br>mlir::Operation::create(...)<br>mlir::detail::OperandStorage::OperandStorage(...) ← crashed here, reading NULL<br>
An optimization pass called FoldMultiplyIntoSDPAScale (it merges a multiply into the "scaled dot-product attention" step of a transformer model) hit a null pointer while rewriting the graph. That code lives in /System/Library/Frameworks/MetalPerformanceShadersGraph.framework — an OS framework, version 7.0.75.2 from the beta, not part of Logic Pro.

★ Insight ─────────────────────────────────────

  • Stem Splitter is a transformer-based (attention-using) neural network. Before running it on the GPU, macOS JIT-compiles the model through an MLIR pipeline — the same compiler infrastructure LLVM uses — and crashes in an optimizer pass, before your audio is even fully processed.
  • The "Translation fault, byte read at 0x0" plus far: 0x0 in the register dump confirms it read through a null pointer rather than, say, corrupted memory — typical of a compiler-pass bug on a graph shape it didn't expect.
  • Because the pass is new-ish (SDPA folding is a recent MPSGraph optimization), this smells like a fresh regression in the macOS 27 beta.
    ─────────────────────────────────────────────────

What you can do​

  1. File it with Apple via Feedback Assistant — attach this .ips file. Beta-OS crashes in system frameworks are exactly what the beta program is for, and this one has a clean, symbolicated stack pointing at their code.
  2. Wait for the next macOS 27 beta / Logic update — the fix has to come from Apple either way (an MPSGraph fix in the OS, or Logic/CoreML sidestepping the bad pass).
  3. Workaround in the meantime: run Stem Splitter on a machine with a release macOS (26.x), or split the stems with a third-party tool (e.g. Demucs-based apps) until the beta is fixed. There's no Logic preference to force the stem model onto CPU, unfortunately.
  4. Rolling back to the stable macOS release is the guaranteed fix if stem splitting is blocking your work.
Nothing in the report implicates your project file, plugins, or hardware — the MIDI control-surface plugins and Audio Units in the binary list were all idle, parked threads.
Click to expand...
Great explanation!

Agree that we need to see if DB4 fixes this, or... if it requires something additional on the Logic side as well, as the DB3 release notes make mention that developers (perhaps here) need to enable a new entitlement??? I don't know - but on my two machines with DB3 - the crash occurs when selecting "Stem Splitter" - it begins to show the UI, then crashes, but the actual stem separation should occur after the UI opens, and the user selects how many "splits" the stem splitter should actually create, so I don't know if the neural network code has actually already been invoked??? Of course, the crash log would indicate it had been, but I don't know why - because the offered splits stem splits is not audio file context sensitive - it will offer to split out drums whether there are drums/percussion in the track to stem split.

This function *does* work as expected on the iPad version of Logic Pro, on DB3 of iPadOS.

With any luck, we'll know today, if there's a DB4 and if this crash goes away or not.

This being the sentence in the release notes that made me wonder...

  • Access to the Neural engine when your app is in the background requires the new entitlement: “com.apple.developer.background-tasks.continued-processing.inference”. (179282606)
 
Michaelhuisman said:
Great explanation!

Agree that we need to see if DB4 fixes this, or... if it requires something additional on the Logic side as well, as the DB3 release notes make mention that developers (perhaps here) need to enable a new entitlement??? I don't know - but on my two machines with DB3 - the crash occurs when selecting "Stem Splitter" - it begins to show the UI, then crashes, but the actual stem separation should occur after the UI opens, and the user selects how many "splits" the stem splitter should actually create, so I don't know if the neural network code has actually already been invoked??? Of course, the crash log would indicate it had been, but I don't know why - because the offered splits stem splits is not audio file context sensitive - it will offer to split out drums whether there are drums/percussion in the track to stem split.

This function *does* work as expected on the iPad version of Logic Pro, on DB3 of iPadOS.

With any luck, we'll know today, if there's a DB4 and if this crash goes away or not.

This being the sentence in the release notes that made me wonder...

  • Access to the Neural engine when your app is in the background requires the new entitlement: “com.apple.developer.background-tasks.continued-processing.inference”. (179282606)
Click to expand...
It works perfectly on Beta 4... yay!!
 
  • Like
Reactions: ProTruckDriver
You must log in or register to reply here.
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.
Top Bottom
Back