While compiling can be thought of as an optimization, many loss functions would jit just fine by default, or could easily be written to be jitable form the beginning. So we would have faster code and therefore a faster dev cycle if we jit by default and turn off jit selectively for models that present some special difficulty.
While compiling can be thought of as an optimization, many loss functions would jit just fine by default, or could easily be written to be jitable form the beginning. So we would have faster code and therefore a faster dev cycle if we jit by default and turn off jit selectively for models that present some special difficulty.