This came up in context of numpy/numpy#30494.

Currently the raw allocator (PyMem_RawMalloc and friends) uses the system allocator because it should be possible to call these APIs without an active thread state. pymalloc requires an active thread state hence that cannot be used, however on free-threading we can instead use mimalloc instead of directly calling the system allocator.

Using mimalloc instead of system allocator I see much less contention while freeing the memory on macos, see the benchmark on the numpy issue on how to reproduce the issue.

Linked PRs

• gh-144914: use mimalloc for raw allocations on free-threading #144916