I set torch, random, and numpy seeds, and my small fine-tune still moves run to run on a T4. Same recipe, same data — the weights at the same step differ between runs. For those of you who actually got reproducible fine-tunes: which flags mattered in practice — math attention backend only, a cuBLAS workspace config, cudnn.deterministic plus benchmark=False, or torch.use_deterministic_algorithms as a diagnostic? And what slowdown did you accept for it? I'm trying to find out which of these is the one that actually closes the gap, rather than turning them all on and hoping.