solprop.backends.mlx_gpu

The mlx backend: single-precision propagation on Apple-silicon GPUs (SEP 33).

Registered as "mlx" and installed with the mlx extra on macOS on Apple silicon. It runs precision="single" problems only, validated against solprop.backends.numpy_ref.single, its oracle (ADR-054). The package is named mlx_gpu so that it does not shadow the mlx library inside it.

Importing it sets MLX_ENABLE_TF32=0 when the variable is unset, so that MLX computes float32 matrix products in float32 rather than TF32; MLX reads the setting when its subsystem first initialises, so a process that used MLX before is checked by gemm_probe() and refused if its products are TF32.

Modules

adams	The exponential Adams history on the Metal device, for the mlx backend.
arithmetic	The protected operations of the mlx backend, and the probes that establish them.
backend	The mlx backend: the single-precision oracle's step, on the Metal GPU.
batch	Fixed-schedule lockstep batches of mlx members (SEP 33 section 6.3, Q10).
envelope	The envelope-grid plasma on the Metal device (SEP 33 section 6.3, phase 4).
passes	The passes of one stage on the device, in binary32: the oracle's, in MLX.