Unlike Numba, Triton operators operate on blocks with explicit load and store of blocks. This is what enables analysis to automate coalescing, shared memory management, etc.

I guess Triton is not for you if Numba is fast enough.