On my Framework Desktop (Strix Halo) running CachyOS with ROCm 7.2.2 on gfx1151, several 8-bit Qwen MLX models produced broken output in lemon-mlx-engine, including:
- mlx-community/Qwen3.5-27B-8bit
- mlx-community/Qwen3.6-27B-8bit
- mlx-community/Qwen3-1.7B-8bit
This turned out to be an upstream MLX ROCm issue in the tiled 8-bit QMV path, not a lemon-mlx-engine bug.
Upstream fix: NripeshN/mlx#6
Once that MLX change is merged, this downstream issue should be resolved.
On my Framework Desktop (Strix Halo) running CachyOS with ROCm 7.2.2 on gfx1151, several 8-bit Qwen MLX models produced broken output in lemon-mlx-engine, including:
This turned out to be an upstream MLX ROCm issue in the tiled 8-bit QMV path, not a lemon-mlx-engine bug.
Upstream fix: NripeshN/mlx#6
Once that MLX change is merged, this downstream issue should be resolved.