Ablation on Attention Mask Type

[zhi-dian]

Hi, seems like the draft model only processes token once, different from other block dLMs that denoise several times, making me thinking if it more like an AR model. Have you tried ablation on using causal mask for dflash?