ML compilers in particular go beyond even the level of stability you would expect from numerical programs. Due to how the SIMT model of thread/warp divergence works, the hardware heavily punishes unstable branches. E.g. if you have 32 threads taking a branch then recoalescing on a barrier -- if they all go the same direction then they can go down the execution pipe as a single bundle, but if 1 takes it while 31 don't, then that's 2x the ex-pipe usage by default (and if you have e.g. a computed-branch, performance goes out the window). Consequently, the whole stack is built around the expectation of stable control flow, even to the detriment of performance (from a local perspective).

ML frameworks even take advantage of this to compute, ahead-of-time, how much memory will be used at different points in the program graph, and thereafter schedule memcpy's to make space as necessary. Of course this only works for well-behaved program classes, but e.g. most LLM architectures fit into that category. Interestingly MoE models don't, since they require data-dependent control flow, thus the recent push towards accommodating dynamism in frameworks (like JAX, which until ~recently couldn't handle it at all).