I think this is not possible because of CUDA requirements; you can force deterministic mode but in my experience results will still differ across hardware.

→ View original post on X — @rasbt