MLX 0.31.2-dev — Per-Stream Lock Patch (pre-release)

Pre-release build of MLX with per-stream locking for concurrent Metal inference thread safety.

PR: ml-explore#3247

Install

# Python 3.12
pip install mlx-0.31.2.dev20260312+0358c602-cp312-cp312-macosx_14_0_arm64.whl

# Python 3.14
pip install mlx-0.31.2.dev20260312+0358c602-cp314-cp314-macosx_14_0_arm64.whl

Requires Apple Silicon Mac, macOS 14+. Replaces your existing mlx install. To revert: pip install mlx==0.31.1

What it fixes

Concurrent GPU evaluations on separate streams (e.g. multi-request inference servers) no longer crash with SIGSEGV/SIGABRT in Metal's command encoder lifecycle.

Testing

355/355 requests across 3 model architectures at up to 20 concurrent streams — zero crashes.