10-17-2026, 04:24 PM
Interesting, I hadn't tried that. I ended up in about the same place using MLX. The biggest change for me was simplifying the stack instead of adding one more optimization layer.
7900 XTX / ROCm / somehow still optimistic