i've heard it actually depends on the model / hosting architecture. some are not deterministic at the numeric level because there is so much floating point math going on in distributed fashion across gpus, with unpredictable rounding/syncing across machines