macOS 27 IOGPUFamily changes and whether the panic is fixed

Parent: Mac local LLMs: GPU stability and kernel panics · Published reference · snapshot 2026-10-05

Apple's macOS 27 release notes list no IOGPUFamily, kernel-panic, GPU-watchdog or Metal-hang entry. Their Metal section holds only two sampler clamp-to-edge fixes.

These notes link each claim to its source. A source may be a research report hosted on this site rather than the primary document. A published reference means the content is available; it does not certify independent review or accuracy.Read the editorial policy and follow the sources before relying on a claim.

Facts

Apple's macOS 27 release notes list no IOGPUFamily, kernel-panic, GPU-watchdog or Metal-hang entry. Their Metal section holds only two sampler clamp-to-edge fixes. [source]
The only macOS 27 evidence on the panic is a mitigation that still works on IOGPUFamily 162.11. With wiring on, nobody has reported a result, because testing risks a reboot. [source]
A different macOS 27 failure is documented in oMLX 3706: `mx.synchronize()` in the periodic cache-clear drain fails with `kIOGPUCommandBufferCallbackErrorTimeout`, and after one to three of these the driver rejects all later submissions (`SubmissionsIgnored`, a "penalty box"). It needs a second process actively submitting Metal work. [source]
oMLX 4224's Hang class does not need macOS 27: two of its ten events ran on 26.5.1. [source]
2026-09-14: macOS 27.0 "Golden Gate" ships as build 26A428 (SDK 26A425 in Xcode 27 27A266a). [source]
2026-09-15 17:56: the oMLX 3706 host upgrades from 26.5.1; the first Timeout arrives 2 h 23 m after the first macOS 27 boot, after zero in 156 engine starts since 2026-06-20. [source]
2026-09-16: oMLX commit 5941f68 makes `SubmissionsIgnored` terminate the process so a supervisor restarts it. [source]
2026-09-21: macOS 27.2 beta 2 (26B5091g). 2026-09-28: macOS 27.0.1 (26A434), published with no developer release notes, update text "bug fixes for your Mac". [source]
2026-10-03: oMLX 4224 opens with Hang events on 27.0 and 26.5.1. [source]
oMLX 3706 exposure table (reporter's own probe series, M3 Ultra 512 GB, oMLX 0.5.7, mlx 0.32.0): macOS 26 with co-host serving, 130 completions, 0 faults; macOS 27 with co-host serving, 41 completions, 106 faults; macOS 27 with co-host resident but idle, 1.5 h, 0 faults; macOS 27 solo, 18.6 h, 0 faults. [source]
The first Timeout was absorbed twice before the third tipped the process into SubmissionsIgnored; after a restart a single Timeout tipped it at once, so the driver's error budget is not simply per process lifetime. [source]
A held process in the penalty box keeps about 400 GB wired and still answers `/v1/models` in 3 ms while generating nothing, so only an external generation probe or exit-on-SubmissionsIgnored detects it. [source]
The reporter first blamed a scheduling-priority gap (co-host at `Nice=-5`, `ProcessType=Interactive`), measured it on 09-20, and withdrew it: priority parity changed nothing. [source]
The maintainer could not reproduce the original Timeout on a newer oMLX and MLX and shipped only the exit patch, so the Timeout root cause is open. [source]
A separate open oMLX thread, issue 1835, is titled "macOS 27: long-context inference 10x slower than macOS 26.6 (even after HOST_VM_INFO64 fix)"; its body was not read. [source]
Is macOS 27 required for the concurrent-submission failures? The 3706 reporter says "necessary but not sufficient" for the Timeout class; the 4224 reporter says not required for the Hang class (2 of 10 events on 26.5.1). Different error classes, different hosts, not reconciled. [source]
Is the prepare-count panic fixed on 27? No source says yes; MetalGuard's 0.9.0 notes say 26.4.x and a 26.5 beta had not fixed it (2026-04-25). Silence on 27 is not evidence of a fix. [source]
Does IOGPUFamily 162.x still underflow the prepare count with MLX wiring on? Untested by anyone found. [source]
What changed on macOS 27 so that a second Metal process trips Timeout in the drain? Maintainer could not reproduce on newer builds; Apple silent. [source]
Does 27.0.1 (26A434) or 27.2 change any of this? No report found. [source]
macOS 27.0 shipped on 2026-09-14 as build 26A428 with SDK build 26A425 inside Xcode 27 (27A266a). [source]
macOS 27.2 beta 2 is build 26B5091g dated 2026-09-21, and macOS 27.0.1 is build 26A434 dated 2026-09-28 with no developer release notes. [source]
Apple's update text for macOS 27.0.1 is "This update provides bug fixes for your Mac." [source]
Apple's macOS 27 release notes contain no entry naming IOGPUFamily, a kernel panic, a GPU watchdog or a Metal hang; the Metal section holds only two sampler clamp-to-edge fixes. [source]
Apple's notes refer to an "Apple 10 GPU family" and say arm64e.x1 devices include Mac computers with M6 or later chips. [source]
oMLX issue 3706 (opened 2026-09-16, M3 Ultra 512 GB, oMLX 0.5.7, mlx 0.32.0, mlx-lm 0.31.3, `iogpu.wired_limit_mb` 507904) reports `kIOGPUCommandBufferCallbackErrorTimeout` from `mx.synchronize()` in `_sync_and_clear_cache` on macOS 27.0. [source]
The 3706 host saw zero Timeout or SubmissionsIgnored errors across 156 engine starts from 2026-06-20 on macOS 26.5.1, and the first Timeout 2 h 23 m after its first macOS 27 boot. [source]
The 3706 host recorded 106 faults after the upgrade; every fault hour but two had a banked probe row showing a second Metal-heavy process actively serving. [source]
On the 3706 host, macOS 26 with the second process serving gave 130 completions and 0 faults; macOS 27 with it serving gave 41 completions and 106 faults. [source]
On the 3706 host, macOS 27 with the second process resident but idle gave 1.5 h and 0 faults, and macOS 27 solo gave 18.6 h and 0 faults. [source]
The 3706 reporter concludes macOS 27 is "necessary but not sufficient" and that the second required ingredient is concurrent submission, not co-residence. [source]
The 3706 reporter withdrew a scheduling-priority theory after removing `Nice=-5` and `Interactive` from the co-host changed nothing. [source]
On the 3706 host a Timeout was absorbed twice and the third tipped the process into `SubmissionsIgnored` in the same second; after a restart a single Timeout tipped it at once. [source]
A 3706 engine took its first Timeout about 2.5 minutes after model load, inside its own startup decode smoke test, printed ready anyway, and could not generate 11 minutes later. [source]
The oMLX maintainer could not reproduce the original Timeout on a newer build and committed `5941f68` so that `SubmissionsIgnored` terminates the process. [source]
The 3706 reporter says a plain second `omlx serve` under sustained load beside a large resident model should reproduce the fault. [source]
oMLX issue 3706 links an open thread, issue 1835, titled "macOS 27: long-context inference 10x slower than macOS 26.6 (even after HOST_VM_INFO64 fix)". [source]
Of oMLX 4224's ten Hang events, eight ran on macOS 27.0 (26A428) and two on macOS 26.5.1 (25F80), so the Hang class does not need macOS 27. [source]
On macOS 27.0 GA (oMLX 4224 host, `iogpu.wired_limit_mb` 506880) 503.6 GB wired with 0.22 GB free ran without a hang or a prepare-count panic, using mlx 0.32.2. [source]
MetalGuard 0.9.0 (2026-04-25) states that macOS 26.4.x and a 26.5 beta had not fixed the prepare-count panic. [source]
No source found reports a prepare-count panic on any macOS 27 build or reports a fixed IOGPUFamily version. [source]

Children

No children recorded.

← the whole tree · 3D view· how to read this page