[davidtai]

Summary

The standalone tests of the fork's process allocation cap (tests/standalone/process_allocation_cap.cpp and process_allocation_cap_metal.cpp, PR #15 / #37) ran in no CI job. This PR runs them in Fork Tests, on the macOS Metal build-cpp job, with no new build: the existing build step builds them, and a new step runs them with ctest. The cap is install-once in each process, so each case runs in its own fresh process (one ctest entry per case).

Running them found three real problems. Two are fixed here. The third (no residency sets on the hosted runner) is handled by David's ruling: skip by capability (finding 4).

Changes

• tests/CMakeLists.txt (Metal builds only): executables process_allocation_cap (CPU ledger, with -UNDEBUG because it checks with assert) and process_allocation_cap_metal (links objc for the method interception). One ctest entry for the CPU test and one for each of the 14 Metal cases.
• .github/workflows/fork-tests.yml, job build-cpp: step "Run the standalone allocation cap tests" after the C++ CPU pass: ctest --test-dir build -R '^process_allocation_cap' --output-on-failure --no-tests=error. It writes coverage profiles like the other steps.
• tests/standalone/process_allocation_cap_metal.cpp: the fixes below. The case list and the checks on a device with the fixed heap and residency sets do not change.
• fork.yaml: the allocator section lists tests/CMakeLists.txt and describes the ctest entries and the device-dependent checks; "Fork tooling" describes the new step.

Findings, in CI order

1. The Metal test did not compile on main (000c0149, Fork Tests job 120941291355, https://github.com/Layr-Labs/mlx/actions/runs/40117360878): 15 errors. The upstream merge (ml-explore/mlx Fix foreign buffer ownership ml-explore/mlx#4634) made allocator::malloc return an owning Data and removed allocator::free. Nothing built this file, so it rotted unseen. Fix 5e9ba74f: allocator().malloc / allocator().free, which return and free a plain Buffer as the old functions did (the same port the merge made in tests/residency_tests.cpp).

2. The test assumed a fixed heap (5e9ba74f, job 120942414585, https://github.com/Layr-Labs/mlx/actions/runs/40117610968): 6/15 passed. The hosted runner's GPU is "Apple Paravirtual device". On it, the Metal allocator makes no fixed heap (metal/allocator.cpp, is_vm). install, retained, churn, foreign, gpu required heap_bytes > 0. Fixes a7d8548a, db97e909: the cases require a heap exactly when the device is not paravirtual, check the heap charge that matches the device, and in churn expect one page for the small buffer when it is not from the heap. The binary prints the device name and whether residency sets are enabled.

3. Residency sets are disabled on the hosted runner (a7d8548a, job 120943910531, https://github.com/Layr-Labs/mlx/actions/runs/40118179184): the log shows device: Apple Paravirtual device, residency sets: disabled. ResidencySets needs the Metal 3 GPU family, which the paravirtual device does not report, so ResidencySets::insert does nothing. The 4 registration-* cases inject a failure into the creation of a residency set and check the rollback of the allocator. On this device that path does not exist. They now fail at once with residency sets are disabled on this device. (They also fill one set with a page first, so that the injected failure is reached on a device with or without the heap in set 0.)

4. Ruling, David 2026-10-24: skip by capability, like the 4 GiB capability skip. cfd37c5e: each registration-* case checks at run time ResidencySets::enabled(), the flag that the residency code of the allocator uses (Metal 3 GPU family and macOS 15). When it is false, the case prints SKIP <case>: residency sets are disabled on this device and exits with 77. Their ctest entries declare SKIP_RETURN_CODE 77, so ctest reports them as skipped, not passed. On a device with residency sets they run unchanged. No runner name or CI tag is used. c3694e4a: the CI step runs ctest with -V, because ctest does not print the output of a skipped case otherwise; the log now shows the reason.

ecbbe349 merges main (with #44 and #46) into this branch (merge commit, no rebase).

CI on the current head

Head c3694e4a: ALL GREEN. Fork Tests https://github.com/Layr-Labs/mlx/actions/runs/40122460853 (build-cpp job 120956780754: 11 passed, 4 registration-* skipped, each with SKIP <case>: residency sets are disabled on this device and device: Apple Paravirtual device, residency sets: disabled in the log), Build and Test https://github.com/Layr-Labs/mlx/actions/runs/40122460867, Fork Diff Check https://github.com/Layr-Labs/mlx/actions/runs/40122461191.

Earlier, on db97e909 (before the ruling): Fork Tests https://github.com/Layr-Labs/mlx/actions/runs/40119475665, build-cpp 11/15 with the 4 registration-* cases failing with residency sets are disabled on this device. On ecbbe349 (skip, before -V): Fork Tests https://github.com/Layr-Labs/mlx/actions/runs/40121592444 green, 11 passed and 4 skipped, but the skip reason was not in the log.

🤖 Generated with Claude Code