Update: finished the paper, and as I suspected, there's a serious downside in speed and memory consumption. LLaDA model has to process the entire output sequenc

Update: finished the paper, and as I suspected, there's a serious downside in speed and memory consumption. LLaDA model has to process the entire output sequence on every time step - without anything like KV cache. Also, full quadratic attention happens on the entire output sequence on every time step, which makes it unfeasible for sequence length longer than a few thousand tokens.