MLX failing to load on Linux with CUDA 12 backend

zluda_trace logs (tarball/zip file)

zluda_logs.zip

Description

I downloaded the latest ZLUDA preview and installed MLX with the CUDA 12 backend in an Python venv. Looking at the ZLUDA trace, it seems like it gets decently far, but fails on a cuMemAllocManaged call.

I would just use llama.cpp with the standard ROCm backend, but I wanted to play around with some community models which are only currently available in MLX format. Would love to see this working some day!

Steps to reproduce

1. Create a Python venv for MLX
2. Run pip install mlx[cuda12] mlx-lm
3. Attempt to run mlx_lm.chat

ZLUDA version

7-preview.9

Operating System

Gentoo Linux

GPU

AMD Radeon RX 7600 XT